Skip to content
modelread

Loading devices…

Remembered on this device. Multi-GPU: search “2x”, “8x”, “node”.

What's new on the Hub

New and changed models in the index, last 14 days.

2026-11-18

    • lineage: – → finetune:Qwen/Qwen2.5-Coder-32B
    • layer_kinds: – → full
    • layer_kinds: – → full,swa
    • moe: – → 256x top-8
    • remote_code: – → yes
    • distilled_from: – → gemma-4-31b
    • min_hardware_claims: – → ["deployable in environments ranging from high-end phones to laptops and servers"]
    • speed_claims: – → ["runs almost as fast as a 4B-parameter model"]
    • languages_card: – → multilingual
    • strengths: – → ["instruction following", "logical reasoning", "text comprehension", "mathematics", "science", "coding", "tool usage", "long-tail knowledge"]
    • min_hardware_claims: – → ["To effectively process a 1 million token context, users will require approximately 240 GB of total GPU memory."]
    • speed_claims: – → ["On sequences approaching 1M tokens, the system achieves up to a 3\u00d7 speedup compared to standard attention implementations."]
    • known_issues: – → ["If you encounter out-of-memory (OOM) issues, consider reducing the context length to a shorter value, such as 32,768.", "This model supports only non-thinking mode and does not generate `` `` blocks in its output."]
    • languages_card: – → multiple lan
    • strengths: ["long context", "coding", "agentic knowledge work", "multimodality"] → ["long context", "coding", "agentic work", "multimodality"]
    • speed_claims: – → ["high-throughput inference with minimal latency and cost overhead"]
    • native_low_bit_card: – → yes
    • strengths: – → ["agentic applications", "tool calling", "local inference", "fine-tuning"]
    • weaknesses: – → ["agentic coding", "complex multi-step reasoning", "knowledge-heavy tasks"]
    • min_hardware_claims: – → ["At 4-bit precision, it takes about 1 GB of memory, so it fits on any phone or laptop."]
    • known_issues: – → ["It is not recommended for agentic coding or complex multi-step reasoning due to its small size."]
    • languages_card: – → english,arabic,chinese,french,german,italian,japanese,korean,portuguese,spanish,vietnamese,thai
    • languages_card: – → 100+ languag
    • strengths: ["instruction following", "long-tail knowledge", "subjective tasks", "long context", "tool calling"] → ["instruction following", "logical reasoning", "mathematics", "science", "coding", "tool usage", "long-tail knowledge", "subjective tasks"]
    • layer_kinds: – → full
    • moe: – → 128x top-8
    • layer_kinds: – → full,linear
    • lineage: – → finetune:Qwen/Qwen3.5-35B-A3B-Base
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • lineage: – → finetune:google/gemma-4-12B
    • layer_kinds: – → full,swa
    • quant_method: – → fp8
    • layer_kinds: – → mla
    • moe: – → 128x top-4
    • layer_kinds: – → full,linear
    • lineage: – → finetune:Qwen/Qwen3.5-0.8B
    • layer_kinds: – → full,linear
    • lineage: – → finetune:openbmb/MiniCPM-V-4.6
    • layer_kinds: – → full,linear
    • layer_kinds: – → full
    • moe: – → 64x top-6
    • remote_code: – → yes
    • lineage: – → finetune:LiquidAI/LFM3-1.7B-Base
    • layer_kinds: – → conv,full,swa
    • remote_code: – → yes
    • layer_kinds: – → full,linear
    • lineage: – → finetune:Nanbeige/Nanbeige4.3-33B-Base
    • layer_kinds: – → full
    • strengths: – → ["agentic coding", "agentic cowork", "reasoning", "tool calling"]
    • moe: 128x top-0 → 128x top-8
    • unmodeled_arch_fields: – → engram_embed_random_seed,engram_fp8_embedding,engram_isolate_fffd_namespace,engram_layer_head_vocab_sizes,engram_layer_ids,engram_max_ngram_size
    • layer_kinds: – → mla
    • moe: – → 512x top-16
    • distilled_from: – → Hy4 flash
    • languages_card: – → english,chinese
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • layer_kinds: – → full,mamba,none
    • moe: – → 512x top-22
    • remote_code: – → yes
    • strengths: – → ["reasoning", "coding", "tool use", "document understanding", "video understanding", "customization", "domain adaptation", "full-precision evaluation"]
    • weaknesses: – → ["accuracy"]
    • lineage: – → finetune:Qwen/Qwen3.6-35B-A3B
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • remote_code: – → yes
    • min_hardware_claims: – → ["Use a GPU with enough memory for the BF16 weights (approximately 69 GB) plus inference state."]
    • known_issues: – → ["very long generations can sometimes become repetitive", "Small numerical differences between its operations and SGLang's kernels can lead to different generated text"]
    • strengths: ["math", "stem", "instruction following", "tool calling"] → ["math", "stem", "programming", "tool calling"]
    • weaknesses: ["coding", "safety"] → ["coding", "agentic"]
    • native_low_bit_card: – → no
    • distilled_from: – → Inkling 2 Small
    • languages_card: – → english,multilingual
    • layer_kinds: – → mla
    • moe: – → 256x top-8
    • strengths: – → ["long context", "coding", "efficiency", "speculative decoding"]
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • distilled_from: – → MiniCPM-V 4.7
    • lineage: – → finetune:Qwen/Qwen3.8-27B
    • layer_kinds: – → full,linear
    • distilled_from: – → Qwen/Qwen3.8-27B
    • layer_kinds: – → full
    • native_low_bit_card: – → yes
    • strengths: – → ["coding", "mathematics", "long context", "tool use", "agentic tasks"]
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • strengths: – → ["agentic orchestration", "verification", "tool use", "scientific reasoning", "instruction following"]
    • layer_kinds: – → full,linear
    • strengths: – → ["ocr accuracy", "multilingual", "layout analysis", "table extraction"]
    • weaknesses: – → ["table geometry"]
    • min_hardware_claims: – → ["throughput of 4.5 pages/s on one H100 at 192 DPI"]
    • speed_claims: – → ["4.5 pages/s", "throughput of 4.5 pages/s on one H100 at 192 DPI, 1.7x Surya 2 with MTP"]
    • known_issues: – → ["row and column geometry needs a Surya 2 checkpoint"]
    • quant_method: – → modelopt
    • lineage: – → merge:autotrust/Qwen3.8-Flash-Next-E384-FP4-Local
    • layer_kinds: – → full,linear
    • moe: – → 384x top-10
    • distilled_from: – → nvidia/Qwen3.8-Flash-Next-FP4-Local
    • known_issues: – → ["CVSS arithmetic is still not reliable", "exact agreement with the reference implementation was about 4.5% for CVSS3 and 4.2% for CVSS4"]
    • quant_method: – → fp8
    • unmodeled_arch_fields: – → hc_mult,compress_ratios
    • layer_kinds: – → mla
    • moe: – → 512x top-10
    • native_low_bit_card: – → yes
    • languages_card: – → multilingual
    • quant_method: – → modelopt
    • lineage: – → quantized:nvidia/Qwen3.8-Flash-Next-FP4-Local
    • layer_kinds: – → full,linear
    • moe: – → 384x top-10
    • pruning: – → reap
    • pruning_calibration: – → c-eval,mmlu,gpqa,aime,humaneval,mmmu
    • unmodeled_arch_fields: – → hc_mult
    • layer_kinds: – → mla
    • moe: – → 64x top-4
    • remote_code: – → yes
    • strengths: – → ["agent", "long context"]
    • languages_card: – → chinese,english
    • min_hardware_claims: ["deployable in environments ranging from high-end phones to laptops and servers", "Smaller models are specifically designed for efficient local execution on laptops and mobile devices"] → ["deployable in environments ranging from high-end phones to laptops and servers"]
    • speed_claims: ["runs much faster than its 26B total might suggest", "runs almost as fast as a 4B-parameter model"] → ["runs almost as fast as a 4B-parameter model"]
    • languages_card: multilingual,out-of-the-b → multilingual
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • strengths: – → ["tool use", "coding"]
    • weaknesses: – → ["generalization"]
    • layer_kinds: – → mla
    • moe: – → 512x top-8
    • remote_code: – → yes
    • quant_method: – → compressed-tensors
    • lineage: – → quantized:meta-llama/Llama-3.3-70B-Instruct
    • layer_kinds: – → full
    • speed_claims: – → ["increasing matrix-multiply compute throughput (by approximately 2x)"]
    • languages_card: – → english,german,french,italian,portuguese,hindi,spanish,thai
    • strengths: ["coding", "mathematics", "instruction following", "long text generation", "structured data understanding", "structured output generation", "system prompt resilience", "long context"] → ["coding", "mathematics", "instruction following", "long text generation", "structured data understanding", "structured output generation", "system prompt resilience", "role-play"]
    • strengths: ["instruction following", "logical reasoning", "mathematics", "science", "coding", "tool usage", "long-tail knowledge", "subjective tasks"] → ["instruction following", "long-tail knowledge", "subjective tasks", "long context", "tool calling"]
    • distilled_from: – → gpt-5
    • min_hardware_claims: – → ["deployable in environments ranging from high-end phones to laptops and servers"]
    • speed_claims: – → ["runs almost as fast as a 4B-parameter model"]
    • languages_card: – → multilingual
    • strengths: ["visual agent", "visual coding", "spatial perception", "long context", "multimodal reasoning", "ocr"] → ["visual agent", "visual coding", "spatial perception", "long context", "multimodal reasoning", "visual recognition", "ocr"]
    • lineage: – → quantized:Qwen/Qwen3-VL-4B-Instruct
    • layer_kinds: – → full
    • strengths: ["visual agent", "visual coding", "spatial perception", "long context", "multimodal reasoning", "visual recognition", "ocr"] → ["visual agent", "visual coding", "spatial perception", "long context", "multimodal reasoning", "ocr"]
    • strengths: ["instruction following", "long-tail knowledge", "subjective tasks", "long context", "tool calling"] → ["instruction following", "logical reasoning", "mathematics", "science", "coding", "tool usage", "long-tail knowledge", "subjective tasks"]
    • lineage: – → quantized:Qwen/Qwen3.6-27B
    • layer_kinds: – → full,linear
    • strengths: – → ["agentic coding", "thinking preservation", "tool calling"]
    • weaknesses: – → ["memory usage", "context length trade-off"]
    • speed_claims: – → ["Inference efficiency and throughput vary significantly across frameworks."]
    • known_issues: – → ["Inference efficiency and throughput vary significantly across frameworks"]
    • lineage: – → finetune:meta-llama/Llama-3.1-70B
    • distilled_from: – → Llama 3.3
    • strengths: – → ["multilingual dialogue", "tool calling"]
    • weaknesses: – → ["languages beyond supported"]
    • languages_card: – → english,german,french,italian,portuguese,hindi,spanish,thai
    • speed_claims: – → ["1100 tokens per second"]
    • languages_card: – → multilingual
    • distilled_from: – → gpt-5
    • min_hardware_claims: – → ["deployable in environments ranging from high-end phones to laptops and servers"]
    • speed_claims: – → ["runs almost as fast as a 4B-parameter model"]
    • languages_card: – → multilingual
    • strengths: – → ["agentic coding", "thinking preservation"]
    • languages_card: – → english,chinese
    • strengths: ["coding", "mathematics", "instruction following", "long text generation", "structured data understanding", "structured output generation", "system prompt resilience", "role-play"] → ["coding", "mathematics", "instruction following", "long text generation", "structured data understanding", "structured output generation", "system prompt resilience", "long context"]
    • strengths: ["visual agent", "visual coding", "spatial perception", "long context", "multimodal reasoning", "ocr"] → ["visual agent", "visual coding", "spatial perception", "long context", "multimodal reasoning", "visual recognition", "ocr"]
    • distilled_from: – → gpt-5
    • min_hardware_claims: – → ["deployable in environments ranging from high-end phones to laptops and servers"]
    • speed_claims: – → ["runs almost as fast as a 4B-parameter model"]
    • languages_card: – → multilingual
    • strengths: – → ["coding", "mathematics", "instruction following", "long text generation", "structured data understanding", "structured output generation", "system prompt resilience", "role-play"]
    • languages_card: – → chinese,english,french,spanish,portuguese,german,italian,russian,japanese,korean,vietnamese,thai
    • strengths: – → ["visual agent", "visual coding", "spatial perception", "long context", "multimodal reasoning", "ocr"]
    • weaknesses: – → ["transformers compatibility"]
    • speed_claims: – → ["Performance metrics are nearly identical to the original BF16 model"]
    • known_issues: – → ["Transformers does not support loading these weights directly"]
    • languages_card: – → ocr supports,quote: "supp
    • strengths: – → ["coding", "mathematics", "instruction following", "long text generation", "structured data understanding", "structured output generation", "system prompt resilience", "long context"]
    • weaknesses: – → ["short text performance with YaRN"]
    • known_issues: – → ["vLLM only supports static YARN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts"]
    • languages_card: – → chinese,english,french,spanish,portuguese,german,italian,russian,japanese,korean,vietnamese,thai
    • strengths: – → ["instruction following", "long-tail knowledge", "subjective tasks", "long context", "tool calling"]
    • min_hardware_claims: – → ["consider reducing the context length to a shorter value, such as 32,768"]
    • speed_claims: – → ["For convenience and performance, we have provided fp8-quantized model checkpoint"]
    • known_issues: – → ["out-of-memory (OOM) issues", "language mixing", "slight decrease in model performance"]
    • languages_card: – → multiple lan
    • min_hardware_claims: – → ["NVIDIA GPU (CUDA)", "Apple Silicon (Metal)", "Intel GPUs (SYCL)"]
    • lineage: – → quantized:Qwen/Qwen3.5-9B
    • layer_kinds: – → full,linear
    • moe: 128x top-0 → 128x top-8
    • lineage: – → quantized:google/gemma-4-26B-A4B-it
    • layer_kinds: – → full,swa
    • moe: – → 128x top-8
    • lineage: – → finetune:Qwen/Qwen3.5-9B
    • layer_kinds: – → full,linear
    • lineage: – → finetune:Qwen/Qwen3.5-9B
    • layer_kinds: – → full,linear
    • layer_kinds: – → full
    • moe: – → 128x top-8
    • strengths: – → ["reasoning", "instruction following", "agent capabilities", "multilingual support", "thinking mode switching"]
    • languages_card: – → 100+ languag
    • strengths: – → ["reasoning", "instruction following", "agent capabilities", "multilingual support"]
    • languages_card: – → 100+ languag
    • strengths: – → ["visual agent", "visual coding", "spatial perception", "long context", "multimodal reasoning", "visual recognition", "ocr"]
    • strengths: – → ["agentic coding", "thinking preservation"]
    • layer_kinds: – → full,swa
    • moe: – → 128x top-0
    • quant_method: – → compressed-tensors
    • lineage: – → quantized:google/gemma-4-12B-it-qat-q4_0-unquantized
    • layer_kinds: – → full,swa
    • quant_method: – → fp8
    • lineage: – → quantized:Qwen/Qwen3-14B
    • layer_kinds: – → full
    • layer_kinds: – → full
    • moe: – → 256x top-16
    • remote_code: – → yes
    • lineage: – → finetune:Qwen/Qwen2.5-32B
    • layer_kinds: – → full
    • layer_kinds: – → full,linear
    • moe: – → 512x top-10
    • layer_kinds: – → full,swa
    • moe: – → 256x top-8
    • remote_code: – → yes
    • layer_kinds: – → mla
    • moe: – → 256x top-8
    • quant_method: – → modelopt
    • lineage: – → quantized:Qwen/Qwen3.8-27B
    • layer_kinds: – → full,linear
    • quant_method: – → awq
    • lineage: – → quantized:Qwen/Qwen2.5-Coder-7B-Instruct
    • layer_kinds: – → full
    • layer_kinds: – → full,linear
    • lineage: – → quantized:Qwen/Qwen3-4B
    • layer_kinds: – → full
    • lineage: – → quantized:Qwen/Qwen2.5-7B-Instruct
    • layer_kinds: – → full
    • lineage: – → quantized:Qwen/Qwen3.8-Flash-Next
    • custom_runtime: – → strata
    • layer_kinds: – →
    • layer_kinds: – → mla
    • moe: – → 896x top-16
    • remote_code: – → yes
    • distilled_from: – → Kimi K3
    • strengths: – → ["long context", "coding", "agentic knowledge work", "multimodality"]
    • layer_kinds: – → full,linear
    • quant_method: – → awq
    • lineage: – → quantized:Qwen/Qwen2.5-VL-7B-Instruct
    • layer_kinds: – → full
    • quant_method: – → awq
    • lineage: – → quantized:Qwen/Qwen2.5-Coder-32B-Instruct
    • layer_kinds: – → full
    • quant_method: – → awq
    • lineage: – → quantized:Qwen/Qwen2.5-VL-3B-Instruct
    • layer_kinds: – → full
    • lineage: – → quantized:Qwen/Qwen3-8B
    • layer_kinds: – → full
    • lineage: – → finetune:LiquidAI/LFM2.5-24B-A2B
    • layer_kinds: – →
    • lineage: – → finetune:Jackrong/Qwopus3.8-27B-Flash
    • quant_method: – → modelopt
    • lineage: – → quantized:Qwen/Qwen2.5-VL-7B-Instruct
    • layer_kinds: – → full
    • layer_kinds: – → full,mamba,none
    • moe: – → 512x top-22
    • remote_code: – → yes
    • lineage: – → quantized:nerkyor/Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2
    • layer_kinds: – → full,linear
    • layer_kinds: – → full,linear
    • quant_method: – → fp8
    • lineage: – → quantized:Qwen/Qwen3.5-35B-A3B
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • quant_method: – → fp8
    • layer_kinds: – → full
    • moe: – → 128x top-8
    • quant_method: – → modelopt
    • layer_kinds: – → full,mamba,none
    • moe: – → 128x top-6
    • layer_kinds: – → full,swa
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • lineage: – → finetune:zai-org/GLM-5.3-BF16
    • layer_kinds: – → mla
    • moe: – → 256x top-8
    • layer_kinds: – → full,mamba,none
    • moe: – → 128x top-6
    • layer_kinds: – → full
    • moe: – → 192x top-8
    • quant_method: – → awq
    • lineage: – → quantized:Qwen/Qwen2.5-32B-Instruct
    • layer_kinds: – → full
    • native_low_bit: – → likely
    • lineage: – → finetune:Qwen/Qwen3.6-27B
    • layer_kinds: – → full,linear
    • lineage: – → finetune:Qwen/Qwen2-VL-2B
    • layer_kinds: – → full
    • native_low_bit: – → likely
    • lineage: – → finetune:Qwen/Qwen3.6-27B
    • layer_kinds: – → full,linear
    • layer_kinds: – → full
    • moe: – → 256x top-6
    • quant_method: – → fp8
    • layer_kinds: – → full
    • moe: – → 256x top-8
    • remote_code: – → yes
    • layer_kinds: – → full,mamba,none
    • moe: – → 512x top-22
    • quant_method: – → awq
    • lineage: – → quantized:Qwen/Qwen2.5-14B-Instruct
    • layer_kinds: – → full
    • quant_method: – → awq
    • lineage: – → quantized:Qwen/Qwen3-8B
    • layer_kinds: – → full
    • quant_method: – → modelopt
    • lineage: – → quantized:google/gemma-4-31B-it
    • layer_kinds: – → full,swa
    • lineage: – → finetune:Qwen/Qwen3-30B-A3B-Base
    • layer_kinds: – → full
    • moe: – → 128x top-8
    • layer_kinds: – → full

2026-11-17

    • quant_method: – → modelopt
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • quant_method: – → modelopt
    • layer_kinds: – → full,linear
    • layer_kinds: – → full
    • layer_kinds: – → full
    • layer_kinds: – → mla
    • moe: – → 256x top-8
    • lineage: – → quantized:google/gemma-4-12B-it-qat-q4_0-unquantized
    • layer_kinds: – → full,swa
    • quant_method: – → fp8
    • lineage: – → quantized:Qwen/Qwen3.6-27B
    • layer_kinds: – → full,linear
    • lineage: – → quantized:google/gemma-4-E4B-it-qat-q4_0-unquantized
    • layer_kinds: – → full,swa
    • quant_method: – → awq
    • lineage: – → quantized:Qwen/Qwen2.5-3B-Instruct
    • layer_kinds: – → full
    • layer_kinds: – → full,linear
    • moe: – → 512x top-10
    • layer_kinds: – → full,linear
    • layer_kinds: – → full,mamba,none
    • moe: – → 128x top-6
    • remote_code: – → yes
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • quant_method: – → fp8
    • lineage: – → quantized:Qwen/Qwen3-VL-8B-Instruct
    • layer_kinds: – → full
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • quant_method: – → awq
    • lineage: – → quantized:Qwen/Qwen2.5-7B-Instruct
    • layer_kinds: – → full
    • layer_kinds: – → full
    • moe: – → 256x top-6
    • lineage: – → finetune:Qwen/Qwen2.5-Coder-7B
    • layer_kinds: – → full
    • quant_method: – → fp8
    • lineage: – → quantized:Qwen/Qwen3-4B-Instruct-2507
    • layer_kinds: – → full
    • lineage: – → finetune:Qwen/Qwen2.5-Coder-14B
    • layer_kinds: – → full
    • quant_method: – → modelopt
    • lineage: – → quantized:google/gemma-4-26B-A4B-it
    • layer_kinds: – → full,swa
    • moe: – → 128x top-0
    • quant_method: – → awq
    • lineage: – → quantized:Qwen/Qwen3-14B
    • layer_kinds: – → full
    • layer_kinds: – → full
    • lineage: – → finetune:Qwen/Qwen2.5-3B
    • layer_kinds: – → full
    • lineage: – → finetune:nvidia/NVIDIA-Nemotron-Nano-9B-v2
    • layer_kinds: – → full,mamba,none
    • remote_code: – → yes
    • lineage: – → quantized:Qwen/Qwen3-VL-8B-Instruct
    • layer_kinds: – → full
    • layer_kinds: – → full
    • quant_method: – → fp8
    • lineage: – → quantized:Qwen/Qwen3.6-35B-A3B
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • quant_method: – → fp8
    • lineage: – → quantized:Qwen/Qwen3.8-27B
    • layer_kinds: – → full,linear
    • lineage: – → quantized:Qwen/Qwen3-0.6B
    • min_hardware_claims: – → ["Apple M2 Pro"]
    • known_issues: – → ["CLINC150 is not recommended in this format", "the trained heads cannot be read", "the label-logits readout and the trained head often disagree on this specialist"]
    • languages_card: – → english,portuguese
    • layer_kinds: – → full
    • layer_kinds: – → full
    • custom_runtime: – → ctranslate2
    • layer_kinds: – →
    • lineage: – → quantized:Qwen/Qwen3-ASR-1.7B
    • custom_runtime: – → transcribe.cpp
    • layer_kinds: – →
    • layer_kinds: – → full
    • layer_kinds: – → full
    • lineage: – → finetune:ResembleAI/chatterbox
    • custom_runtime: – → chatterbox
    • layer_kinds: – → full
    • remote_code: – → yes
    • layer_kinds: – → full
    • layer_kinds: – → full
    • layer_kinds: – → full
    • layer_kinds: – → full
    • layer_kinds: – → full
    • lineage: – → finetune:Qwen/Qwen2.5-7B
    • layer_kinds: – → full
    • strengths: – → ["coding", "mathematics", "instruction following", "long text generation", "structured data understanding", "structured output generation", "system prompt resilience", "long context"]
    • weaknesses: – → ["short text performance with YaRN"]
    • known_issues: – → ["vLLM only supports static YARN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts"]
    • languages_card: – → chinese,english,french,spanish,portuguese,german,italian,russian,japanese,korean,vietnamese,thai
    • custom_runtime: – → whisperkit
    • layer_kinds: – →
    • quant_method: – → fp8
    • lineage: – → finetune:deepseek-ai/DeepSeek-V3.2-Exp-Base
    • layer_kinds: – → mla
    • moe: – → 256x top-8
    • strengths: – → ["math olympiad", "reasoning", "agentic task synthesis"]
    • weaknesses: – → ["tool calling", "output parsing"]
    • layer_kinds: – → full,swa
    • remote_code: – → yes
    • layer_kinds: – → full
    • layer_kinds: – → full
    • lineage: – → finetune:Qwen/Qwen3-8B-Base
    • layer_kinds: – → full
    • layer_kinds: – → full
    • lineage: – → quantized:MiniMaxAI/MiniMax-H3-Image
    • custom_runtime: – → minimax-h3
    • lineage: – → quantized:moondream/parakeet-ultra
    • custom_runtime: – → transcribe.cpp
    • layer_kinds: – →
    • layer_kinds: – → full
    • strengths: – → ["generalization", "speech recognition", "speech translation"]
    • weaknesses: – → ["high-risk domains", "classification", "consent"]
    • languages_card: – → english,french,multilingual
    • layer_kinds: – → full
    • layer_kinds: – → full
    • layer_kinds: – → full,linear
    • layer_kinds: – → full
    • remote_code: – → yes
    • layer_kinds: – →
    • remote_code: – → yes
    • lineage: – → finetune:moonshine-ai/moonshine-streaming-tiny
    • layer_kinds: – → full
    • strengths: – → ["low latency", "small size", "confidence scoring"]
    • weaknesses: – → ["accuracy in noise", "rewrites"]
    • min_hardware_claims: – → ["It runs about twice as fast as real time on one CPU thread."]
    • speed_claims: – → ["It runs about twice as fast as real time on one CPU thread."]
    • layer_kinds: – → full
    • lineage: – → finetune:Qwen/Qwen2.5-1.5B
    • layer_kinds: – → full
    • strengths: – → ["coding", "mathematics", "instruction following", "long text generation", "structured data understanding", "structured output generation", "system prompt resilience", "role-play"]
    • languages_card: – → chinese,english,french,spanish,portuguese,german,italian,russian,japanese,korean,vietnamese,thai
    • quant_method: – → modelopt
    • lineage: – → quantized:Qwen/Qwen3.6-35B-A3B
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • layer_kinds: – → full,linear
    • lineage: – → quantized:PrunaAI/P-H3-Image
    • custom_runtime: – → diffusion-single-file
    • layer_kinds: – → full
    • distilled_from: – → FlagEmbedding
    • languages_card: – → english,chinese,multilingual
    • lineage: – → finetune:Qwen/Qwen2.5-0.5B
    • layer_kinds: – → full
    • strengths: – → ["coding", "mathematics", "instruction following", "long text generation", "structured data understanding", "structured output generation", "system prompt resilience", "role-play"]
    • languages_card: – → chinese,english,french,spanish,portuguese,german,italian,russian,japanese,korean,vietnamese,thai
    • layer_kinds: – → full
    • lineage: – → finetune:Qwen/Qwen3-8B-Base
    • layer_kinds: – → full
    • layer_kinds: – → full
    • lineage: – → finetune:Qwen/Qwen3.5-4B
    • layer_kinds: – → full,linear
    • layer_kinds: – →
    • lineage: – → finetune:Qwen/Qwen3-14B-Base
    • layer_kinds: – → full
    • lineage: – → finetune:Qwen/Qwen3-4B-Base
    • layer_kinds: – → full
    • custom_runtime: – → chatterbox
    • distilled_from: – → Chatterbox Multilingual V3
    • strengths: – → ["speaker similarity", "hallucination reduction", "emotion control", "watermarking", "benchmarking"]
    • weaknesses: – → ["language transfer"]
    • speed_claims: – → ["sub 200ms"]
    • known_issues: – → ["hallucinations", "repetition", "off-prompt speech"]
    • lineage: – → finetune:Qwen/Qwen3.5-9B
    • layer_kinds: – → full,linear
    • strengths: – → ["web agent", "screenshot-only", "multi-site", "reasoning"]
    • min_hardware_claims: – → ["single GPU (TP=1)"]
    • languages_card: – → korean,english
    • layer_kinds: – → full
    • remote_code: – → yes
    • distilled_from: – → KLM 1.6
    • min_hardware_claims: – → ["KLM 1.6 runs on a CPU as well as on any GPU without extra packages beyond Transformers."]
    • speed_claims: – → ["fast_generate decodes many prompts at once with fused Triton kernels and CUDA graphs"]
    • known_issues: – → ["The model's knowledge is thin, and the facts it states are often wrong.", "It only works in English and has a 2,048-token context window.", "It has no safety alignment, so add your own filtering before putting it in front of users."]
    • distilled_from: – → FLUX.1 [schnell]
    • min_hardware_claims: – → ["save some VRAM by offloading the model to CPU. Remove this if you have enough GPU power"]
    • speed_claims: – → ["can generate high-quality images in only 1 to 4 steps."]
    • known_issues: – → ["This model is not intended or able to provide factual information.", "As a statistical model this checkpoint might amplify existing societal biases.", "The model may fail to generate output that matches the prompts.", "Prompt following is heavily influenced by the prompting-style."]
    • languages_card: – → english
    • layer_kinds: – →
    • strengths: – → ["voice cloning", "cross-language voice cloning", "emotion and style transfer", "audio quality"]
    • weaknesses: – → ["training data"]
    • languages_card: – → english,spanish,french,german,italian,portuguese,polish,turkish,russian,dutch,czech,arabic
    • distilled_from: – → FLUX.1 [pro]
    • min_hardware_claims: – → ["save some VRAM by offloading the model to CPU"]
    • speed_claims: – → ["more efficient"]
    • known_issues: – → ["amplify existing societal biases", "fail to generate output that matches the prompts", "prompt following is heavily influenced by the prompting-style"]
    • lineage: – → finetune:Qwen/Qwen3-1.7B-Base
    • layer_kinds: – → full
    • weaknesses: – → ["presence_penalty side effects"]
    • known_issues: – → ["endless repetitions"]
    • languages_card: – → 100+ languag,quote: suppo
    • lineage: – → finetune:Qwen/Qwen3-0.6B-Base
    • layer_kinds: – → full
    • strengths: – → ["multilingual", "long context", "text retrieval", "code retrieval", "instruction following", "flexible dimensions"]
    • weaknesses: – → ["instruction dependency"]
    • languages_card: – → 100+ languag,quote: "supp
    • layer_kinds: – → full
    • strengths: – → ["visual agent", "visual coding", "spatial perception", "long context", "multimodal reasoning", "visual recognition", "ocr"]
    • languages_card: – → english,chinese,japanese,korean,french,spanish,german,russian,arabic,portuguese,italian,dutch
    • lineage: – → finetune:nvidia/Nemotron-3.5-Parse
    • layer_kinds: – → full
    • remote_code: – → yes
    • distilled_from: – → nvidia/Nemotron-3.5-Parse
    • min_hardware_claims: – → ["one NVIDIA GPU", "about 8 GB of an H100"]
    • speed_claims: – → ["1.5 s per page on one H100", "19,176 pages an hour with 16 pages in flight"]
    • layer_kinds: – → full
    • moe: – → 128x top-8
    • strengths: – → ["agentic coding", "agentic browser-use", "long context", "repository-scale understanding"]
    • weaknesses: – → ["thinking mode"]
    • known_issues: – → ["This model supports only non-thinking mode and does not generate `` `` blocks in its output."]
    • lineage: – → finetune:Qwen/Qwen3-4B-Base
    • layer_kinds: – → full
    • lineage: – → finetune:nvidia/Cosmos3-Edge
    • layer_kinds: – →
    • lineage: – → finetune:Qwen/Qwen3.5-4B
    • layer_kinds: – → full,linear
    • strengths: – → ["3d scene generation", "general coding"]
    • weaknesses: – → ["one-shot reliability", "variable naming"]
    • known_issues: – → ["duplicate variable names"]
    • languages_card: – → english,quote: "use
    • strengths: – → ["image generation", "transparent images", "text rendering"]
    • lineage: – → finetune:Qwen/Qwen3.5-2B-Base
    • layer_kinds: – → full,linear
    • speed_claims: – → ["high-throughput inference with minimal latency and cost overhead"]
    • languages_card: – → 201 language
    • distilled_from: – → Llama 3 8B
    • strengths: – → ["dialogue", "open source chat", "helpfulness", "safety", "coding"]
    • weaknesses: – → ["non-english languages", "false refusals"]
    • languages_card: – → english
    • lineage: – → quantized:ACE-Step/Ace-Step1.5
    • custom_runtime: – → audio.cpp
    • layer_kinds: – →
    • quant_method: – → fp8
    • unmodeled_arch_fields: – → hc_mult,compress_ratios
    • layer_kinds: – → swa
    • moe: – → 256x top-6
    • layer_kinds: – → full
    • layer_kinds: – →
    • lineage: – → finetune:google/gemma-4-E4B
    • layer_kinds: – → full,swa
    • speed_claims: – → ["you can improve the inference speed by 20-30% with torch.compile", "This technique is slightly slower than the first one, as it requires more function evaluations."]
    • known_issues: – → ["The model does not achieve perfect photorealism", "The model cannot render legible text", "The model struggles with more difficult tasks which involve compositionality, such as rendering an image corresponding to \u201cA red cube on top of a blue sphere\u201d", "Faces and people in general may not be generated properly.", "The autoencoding part of the model is lossy."]
    • layer_kinds: – → full,swa
    • distilled_from: – → MetaEncoder v1
    • lineage: – → quantized:Qwen/Qwen3.8-Flash-Next
    • custom_runtime: – → edge0
    • layer_kinds: – → full,linear
    • moe: – → 512x top-10
    • lineage: – → finetune:google/gemma-4-E2B
    • layer_kinds: – → full,swa
    • languages_card: – → multilingual
    • lineage: – → adapter:Qwen/Qwen-Image-2.1
    • strengths: – → ["anime style", "composition", "nsfw anime image generation", "reference-to-image generation"]
    • weaknesses: – → ["universal superiority", "turbo stability"]
    • speed_claims: – → ["Turbo remains fixed at 8 steps", "20 steps can already give good results for anime illustrations with simple layouts"]
    • known_issues: – → ["Turbo LoRA is still under training and refinement", "Quality, style coverage, and recommended settings may change in later checkpoints", "Do not load the original-Base LoRA on Turbo, and do not load the Turbo LoRA on the original Base", "Do not stack the two LoRAs"]
    • lineage: – → quantized:Qwen/Qwen3.8-Flash-Next
    • custom_runtime: – → edge0
    • layer_kinds: – → full,linear
    • moe: – → 512x top-10
    • lineage: – → finetune:krea/Krea-2-Raw
    • custom_runtime: – → diffusion-single-file
    • native_low_bit_card: – → yes
    • distilled_from: – → krea/Krea-2-Raw
    • lineage: – → finetune:Qwen/Qwen3-0.6B-Base
    • layer_kinds: – → full
    • weaknesses: – → ["presence_penalty side effects"]
    • known_issues: – → ["endless repetitions"]
    • languages_card: – → 100+ languag,quote: suppo
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • distilled_from: – → Qwen3.5 and Gemma4
    • strengths: – → ["coding", "agentic coding", "tool calling"]
    • min_hardware_claims: – → ["2\u00d7 80GB GPUs"]
    • known_issues: – → ["Open-source runtimes implement YaRN statically, which can slightly hurt quality on ordinary-length inputs"]
    • quant_method: – → fp8
    • unmodeled_arch_fields: – → hc_mult
    • layer_kinds: – → mla
    • moe: – → 288x top-8
    • lineage: – → finetune:Qwen/Qwen3.5-0.8B-Base
    • layer_kinds: – → full,linear
    • layer_kinds: – → full
    • moe: – → 64x top-6
    • remote_code: – → yes
    • quant_method: – → mxfp4
    • layer_kinds: – → full,swa
    • moe: – → 32x top-4
    • lineage: – → finetune:google/gemma-4-26B-A4B
    • layer_kinds: – → full,swa
    • moe: – → 128x top-0
    • layer_kinds: – →
    • min_hardware_claims: – → ["The model can decode on a CPU."]
    • speed_claims: – → ["The recorded median CPU decoding time in that run was 18.463 ms using four threads."]
    • known_issues: – → ["Camera framing, omitted details, and numerical edits remain areas for improvement.", "Image input, arbitrary new mesh generation, and numerical rotation commands are outside this checkpoint's current schema."]
    • languages_card: – → english
    • lineage: – → finetune:google/gemma-4-26B-A4B-it
    • layer_kinds: – → full,swa
    • moe: – → 128x top-0
    • remote_code: – → yes
    • distilled_from: – → gpt-5.2-2025-12-11
    • speed_claims: – → ["transcribes up to 1,039 seconds of audio per second", "returns the transcript of a 30-second recording in 0.22 seconds", "speeds up decoding about 2\u20133\u00d7"]
    • layer_kinds: – → full
    • strengths: – → ["multi-functionality", "multi-linguality", "long context", "hybrid retrieval", "benchmark performance"]
    • weaknesses: – → ["fp16 precision", "evaluation methodology"]
    • speed_claims: – → ["Setting use_fp16 to True speeds up computation with a slight performance degradation", "If you don't need such a long length, you can set a smaller value to speed up the encoding process.", "a smaller max length leads to a lower latency"]
    • known_issues: – → ["Setting use_fp16 to True speeds up computation with a slight performance degradation"]
    • languages_card: – → more than 10,13 languages,english
    • quant_method: – → mxfp4
    • layer_kinds: – → full,swa
    • moe: – → 128x top-4
    • native_low_bit_card: – → yes
    • custom_runtime: – → trellis2
    • strengths: – → ["3d generation", "complex topology", "pbr materials", "efficiency", "high resolution", "compact latent space"]
    • weaknesses: – → ["geometric artifacts", "alignment"]
    • min_hardware_claims: – → ["An NVIDIA GPU with at least 24GB of memory is necessary."]
    • speed_claims: – → ["Inference is highly efficient; see table below.", "512\u00b3 | ~3 seconds", "1024\u00b3 | ~17 seconds", "1536\u00b3 | ~60 seconds"]
    • known_issues: – → ["Geometric Artifacts (Small Holes)", "Base Model w/o Alignment"]
    • lineage: – → finetune:Jackrong/DeepSeek-V4-Pro-Qwen3.5-4B
    • layer_kinds: – → full,linear
    • known_issues: – → ["hallucinations", "arithmetic mistakes", "reasoning failures", "unstable behavior on difficult or underspecified tasks"]
    • lineage: – → adapter:MiniMaxAI/MiniMax-H3-Image
    • strengths: – → ["speed", "realtime preview", "image quality"]
    • weaknesses: – → ["stock model limitation", "detail pass artifacts"]
    • min_hardware_claims: – → ["RTX PRO 6000"]
    • speed_claims: – → ["~17.7 s per image", "2 denoising evaluations instead of 6", "2048\u00d72048 renders in ~4.6 s"]
    • layer_kinds: – → full
    • min_hardware_claims: – → ["VRAM requirements"]
    • speed_claims: – → ["4.5x speed-ups"]
    • known_issues: – → ["torch.compile is currently not compatible with the Chunked long-form algorithm or Flash Attention 2"]
    • languages_card: – → cantonese,english,french
    • lineage: – → finetune:openai/whisper-large-v3
    • layer_kinds: – → full
    • speed_claims: – → ["4.5x speed-ups", "way faster"]
    • known_issues: – → ["torch.compile is currently not compatible with the Chunked long-form algorithm or Flash Attention 2", "minor quality degradation"]
    • languages_card: – → english,french
    • lineage: – → finetune:Qwen/Qwen3.5-4B-Base
    • layer_kinds: – → full,linear
    • strengths: – → ["multimodal learning", "architectural efficiency", "reinforcement learning", "global accessibility", "vision-language foundation", "hybrid architecture", "rl generalization", "linguistic coverage"]
    • languages_card: – → 201 language
    • lineage: – → finetune:Qwen/Qwen3-0.6B
    • layer_kinds: – → full
    • speed_claims: – → ["RTF as low as 0.025 (40x faster than real-time)"]
    • languages_card: – → 600+ languag
    • lineage: – → quantized:nreimers/MiniLM-L6-H384-uncased
    • layer_kinds: – → full
    • distilled_from: – → nreimers/MiniLM-L6-H384-uncased
    • lineage: – → finetune:Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B
    • layer_kinds: – → full,linear
    • strengths: – → ["math", "stem", "instruction following", "tool calling"]
    • weaknesses: – → ["coding", "safety"]
    • speed_claims: – → ["36.1% fewer tokens per correct answer, with +0.93 accuracy points", "22.7% fewer tokens per correct answer, with +9.47 accuracy points", "Correct answers / 1M tokens 316.8"]
    • known_issues: – → ["hallucinations", "reasoning errors", "unstable behavior on difficult or underspecified tasks"]
    • lineage: – → finetune:Qwen/Qwen-Image-2.1
    • custom_runtime: – → diffusion-single-file
    • native_low_bit_card: – → yes
    • layer_kinds: – → full,linear
    • moe: – → 256x top-8
    • strengths: – → ["agentic coding", "thinking preservation"]
    • quant_method: – → fp8
    • unmodeled_arch_fields: – → hc_mult
    • layer_kinds: – → mla
    • moe: – → 288x top-8
    • strengths: – → ["multimodal", "efficiency", "coding", "long context"]
    • lineage: – → finetune:Qwen/Qwen3.5-9B-Base
    • layer_kinds: – → full,linear
    • strengths: – → ["vision-language", "multilingualism", "tool calling", "agents", "coding"]
    • languages_card: – → 201 language
    • custom_runtime: – → minimax-h3
    • distilled_from: – → MiniMax H3 Image
    • min_hardware_claims: – → ["NVIDIA with 8 GB VRAM", "System RAM: 24 GB or more", "Disk: about 38 GB free, on an SSD"]
    • known_issues: – → ["Short prompts give a glossy, doll-like face"]
    • lineage: – → finetune:google/gemma-4-31B
    • layer_kinds: – → full,swa
    • min_hardware_claims: – → ["deployable in environments ranging from high-end phones to laptops and servers", "Smaller models are specifically designed for efficient local execution on laptops and mobile devices"]
    • speed_claims: – → ["runs much faster than its 26B total might suggest", "runs almost as fast as a 4B-parameter model"]
    • languages_card: – → multilingual,out-of-the-b
    • lineage: – → finetune:KORMo-VL/KORMo-VL
    • layer_kinds: – → full
    • strengths: – → ["web pretraining scale"]
    • weaknesses: – → ["web task performance"]
    • min_hardware_claims: – → ["single gpu"]
    • languages_card: – → korean,english
    • layer_kinds: – → full,linear
    • distilled_from: – → Qwen3.5
    • min_hardware_claims: – → ["single 80GB GPU", "edge-deployable on mobile devices via its quantized Ornith-1.5-9B-Mobile variant"]
    • known_issues: – → ["Open-source runtimes implement YaRN statically: the same scaling factor is applied to every request regardless of its length, which can slightly hurt quality on ordinary-length inputs.", "Only enable rope_scaling when your workload genuinely needs the longer window, and size factor to match it"]
    • lineage: – → merge:akatz-ai/MiniMax-H3-Character-Swap-LoRA,MiniMaxAI/MiniMax-H3-Image
    • strengths: – → ["character swap", "cross-style swaps", "identity preservation", "pose preservation", "merged checkpoint preference"]
    • weaknesses: – → ["cross-style swaps", "skin tone", "pose", "outfit details", "non-frontal faces", "viewing direction"]
    • min_hardware_claims: – → ["full local run needs ~140 GB VRAM (DiT 66 GB + Qwen3-VL encoder 67 GB)", "only the 66 GB DiT + 2 GB VAE touch your GPU"]
    • speed_claims: – → ["6 sampling steps, CFG 1 (the model is distilled) \u2014 no tuning needed"]
    • known_issues: – → ["cross-style swaps onto non-frontal faces", "skin tone shifts", "pose changes", "outfit details wander", "extreme profile/cross-style combos"]
    • layer_kinds: – → mla
    • moe: – → 896x top-16
    • remote_code: – → yes
    • distilled_from: – → Kimi K3.1 Pro
    • strengths: – → ["coding", "3d creation", "financial research", "factual accuracy", "agentic capabilities"]
    • quant_method: – → fp8
    • unmodeled_arch_fields: – → kv_source_layer_ids,candidate_source_layer_id,hc_mult,engram_layer_ids,engram_num_embeddings,engram_max_ngram_size
    • layer_kinds: – → swa
    • moe: – → 384x top-6
    • lineage: – → finetune:MiniMaxAI/MiniMax-H3
    • custom_runtime: – → diffusion-single-file
    • distilled_from: – → MiniMax H3
    • lineage: – → finetune:yl4579/StyleTTS2-LJSpeech
    • layer_kinds: – → full
    • strengths: – → ["cost efficiency", "speed", "quality"]
    • min_hardware_claims: – → ["Total Training Cost: About $1000 for 1000 hours of A100 80GB vRAM"]
    • speed_claims: – → ["significantly faster and more cost-efficient", "under $1 per million characters of text input"]
    • known_issues: – → ["Any website containing \"kokoro\" in its root domain is NOT owned by and NOT affiliated with this model page or its author"]
    • quant_method: – → fp8
    • layer_kinds: – → mla
    • moe: – → 256x top-8
    • strengths: – → ["coding", "cyber capability", "long-horizon tasks"]
    • custom_runtime: – → minimax-h3
    • distilled_from: – → MiniMax H3
    • languages_card: – → arabic,chinese,english,french,german,italian,japanese,korean,portuguese,russian,spanish
    • custom_runtime: – → pyannote-audio
    • weaknesses: – → ["benchmark performance"]
    • quant_method: – → fp8
    • unmodeled_arch_fields: – → kv_source_layer_ids,candidate_source_layer_id,hc_mult,engram_layer_ids,engram_num_embeddings,engram_max_ngram_size
    • layer_kinds: – → swa
    • moe: – → 384x top-6
    • lineage: – → finetune:meta-llama/Llama-3.1-8B
    • distilled_from: – → Llama 3.1 405B
    • min_hardware_claims: – → ["Training utilized a cumulative of 39.3M GPU hours of computation on H100-80GB (TDP of 700W) type hardware"]
    • speed_claims: – → ["All model versions use Grouped-Query Attention (GQA) for improved inference scalability."]
    • known_issues: – → ["Use in languages beyond those explicitly referenced as supported in this model card", "Use in any manner that violates applicable laws or regulations (including trade compliance laws).", "Use in any other way that is prohibited by the Acceptable Use Policy and Llama 3.1 Community License."]
    • languages_card: – → english,german,french,italian,portuguese,hindi,spanish,thai
    • layer_kinds: – →
    • unmodeled_arch_fields: – → hc_mult
    • layer_kinds: – → mla
    • moe: – → 288x top-8
    • distilled_from: – → Sage S2.1 Flash
    • strengths: – → ["machine learning engineering", "office and professional document process", "tool use and research execution"]
    • lineage: – → quantized:MiniMaxAI/MiniMax-H3-Image
    • custom_runtime: – → minimax-h3
    • layer_kinds: – →
    • min_hardware_claims: – → ["8 GB \u2013 10 GB VRAM", "12 GB VRAM", "12 GB \u2013 16 GB VRAM", "16 GB VRAM", "16 GB \u2013 24 GB VRAM"]
    • speed_claims: – → ["6 distilled sampling steps"]
    • languages_card: – → english
    • lineage: – → finetune:Qwen/Qwen3.8-27B
    • layer_kinds: – → full,linear
    • distilled_from: – → Alibaba's Qwen3.8-27B
    • strengths: – → ["propaganda reduction", "general capability"]
    • weaknesses: – → ["value remediation", "decoding settings", "benchmark comparison"]
    • known_issues: – → ["refusals, deflections, and CCP-framed answers on politically sensitive topics"]
    • layer_kinds: – → full,linear
    • moe: – → 512x top-10
    • custom_runtime: – → pyannote-audio
    • distilled_from: – → pyannote/speaker-diarization-3.0
    • speed_claims: – → ["possibly speed up inference"]
    • known_issues: – → ["use of `onnxruntime`"]
    • languages_card: – → quote:
    • layer_kinds: – → full
    • moe: – → 256x top-12
    • remote_code: – → yes
    • distilled_from: – → Maxwell-V1.0-Flash
    • languages_card: – → thai,lao,myanmar,khmer,shan
    • lineage: – → quantized:Qwen/Qwen3.8-Flash-Next
    • custom_runtime: – → edge0
    • layer_kinds: – → full,linear
    • moe: – → 512x top-10
    • layer_kinds: – →
    • strengths: – → ["agentic tasks", "coding", "data analysis", "long context", "multimodal"]
    • weaknesses: – → ["complex instruction following", "coding", "agent tasks", "web generation"]
    • min_hardware_claims: – → ["With 4-bit weights and an 8-bit KV cache, multimodal inference at a 256K context length requires approximately 15 GiB of GPU memory.", "With 4-bit weight quantization, the model weights occupy approximately 4.9 GB."]
    • speed_claims: – → ["Moving from two to four loops improves LiveCodeBench V6 by 1.04 percentage points and AIDABench data visualization by 1.35 percentage points."]
    • known_issues: – → ["Complex instruction following: The model may misunderstand or miss requirements when given many complex tool definitions or multiple constraints.", "Coding and general agent tasks: Performance can be inconsistent on coding and general agent tasks.", "Complex web generation: Websites with complex 3D scenes and rich animations may not be completed in a single turn."]
    • languages_card: – → english,chinese
    • custom_runtime: – → zonos3
    • layer_kinds: – →
    • distilled_from: – → Wav2Vec2-BERT
    • min_hardware_claims: – → ["runs in real time on phones and laptops", "runs entirely on your own GPU through WebGPU"]
    • speed_claims: – → ["80 ms for ZONOS3-Mini", "457 seconds of audio per second for ZONOS3-Mini"]
    • languages_card: – → english,french,spanish,german,italian,portuguese,dutch,greek,swedish,danish,norwegian,finnish
    • layer_kinds: – →
    • min_hardware_claims: – → ["runs about 70x faster than real time on four CPU threads"]
    • speed_claims: – → ["runs about 70x faster than real time on four CPU threads"]
    • known_issues: – → ["The F2 voice sounds younger than intended.", "Foreign names and unusual spellings are often mispronounced.", "The same text is always read the same way. There is no sampling of prosody.", "Speech is slightly faster than the training recordings at the default speed.", "No streaming."]
    • languages_card: – → english
    • lineage: – → quantized:Qwen/Qwen-Image-2.1
    • layer_kinds: – →
    • min_hardware_claims: – → ["single RTX 5090 (32 GB)", "24 GB card"]
    • known_issues: – → ["tiled decoding adds visible vertical colour streaks", "BF16 pipeline does not fit comfortably beside a full text encoder on a 24 GB card"]
    • lineage: – → finetune:black-forest-labs/flux-3-image-klein
    • custom_runtime: – → diffusion-single-file
    • native_low_bit_card: – → yes
    • distilled_from: – → FLUX 3 Image Klein
    • min_hardware_claims: – → ["int8 files run on 8 GB GPUs such as the RTX 3070", "NVFP4 needs a Blackwell GPU"]
    • speed_claims: – → ["16 steps and no guidance for the distilled checkpoint", "50 steps and CFG 3.5 for the base checkpoint"]
    • lineage: – → finetune:MiniMaxAI/MiniMax-H3-Image
    • custom_runtime: – → diffusion-single-file
    • native_low_bit_card: – → yes
    • distilled_from: – → MiniMax H3 Image
    • quant_method: – → modelopt
    • layer_kinds: – → full
    • moe: – → 384x top-8
    • layer_kinds: – →
    • min_hardware_claims: – → ["NVIDIA GPU (about 12 GB of memory)"]
    • speed_claims: – → ["Native full-duplex at 7.5 Hz", "Every 133 ms the model takes one frame", "produces it in one step from the language model's hidden state"]
    • known_issues: – → ["It is not a reliable source of facts", "does not follow instructions the way an assistant does", "can imitate the voice of a reference clip"]
    • languages_card: – → english,spanish,portuguese,german,french,chinese,japanese,russian,korean
    • custom_runtime: – → diffusion-single-file
    • native_low_bit_card: – → yes
    • layer_kinds: – → full,linear
    • strengths: – → ["coding", "agentic tasks", "vision-language", "thinking control"]
    • weaknesses: – → ["reasoning effort trade-off"]
    • lineage: – → finetune:MiniMaxAI/MiniMax-H3
    • custom_runtime: – → minimax-h3
    • speed_claims: – → ["6 sampling steps", "no classifier-free guidance"]
    • languages_card: – → english
Askwhat should I run?
Browsebest models per device
Devicesall hardware
Connect your agentMCP · API
How the numbers workcalibration