Skip to content
m.recipes
Find a configuration
Sign in
0 models with qualified results

Models

Measured configurations by model. Open a model to compare GPUs on one exact test.

Get your startup command

Choose your operating system and enter a model repository. Launch support is checked separately from performance results.

Reset

All tasks · RTX 4090D · 48 GB

Speeds compare only within one exact test. Recipe availability is checked on each report against the published recipe executor.

No qualified results yet · All tasks · RTX 4090D · 48 GB

445 original reports are kept as evidence but do not qualify for comparison; each report states why. This does not show that the model or GPU cannot run it.

Check model launch supportView original reportsMeasured alternatives for this taskTest your own GPU

Measured alternatives on other GPUs

Measured on different GPUs with the same or smaller listed capacity. These are references for choosing hardware, not performance or compatibility evidence for your selected GPU.

  1. Qwen3.8-27B-NVFP4

    RadixArk · 2 tested configurations · 1 GPU

    • RTX 5090 · 32 GBNVFP4 · 8K → 2K tokens
      Output throughput · includes prefill · 1 request
      70.8tok/s
      Observed GPU occupancy
      27.3GiB
      First token
      633 ms
      Use this recipe · 1 request ↗Evidence report and run availability ↗Same test on other GPUs
      Test details

      Total output tokens divided by the full request window, including prefill and queueing.

      Occupancy includes engine buffers and reserved KV cache. It is not the minimum VRAM required by the model.

      Reports combined counts reports, not inference requests. Request successes and failures are listed in each report.

      Engine
      vllm
      Protocol
      wbench-fixed-http-v1
      Dataset
      synthetic-tokenizer-v1:repeats=2
      Workload
      8192in-2048out
      ConcurrencyMean throughput (tok/s)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
      C170.8229.01 s633 ms27.29405.51bad294c0 ↗
    • 1 other test configurations on this GPU
      • NVFP4 · 32K → 2K tokens
        Output throughput · includes prefill · 1 request
        62.5tok/s
        Observed GPU occupancy
        27.3GiB
        First token
        3.6 s
        Use this recipe · 1 request ↗Evidence report and run availability ↗Same test on other GPUs
        Test details

        Total output tokens divided by the full request window, including prefill and queueing.

        Occupancy includes engine buffers and reserved KV cache. It is not the minimum VRAM required by the model.

        Reports combined counts reports, not inference requests. Request successes and failures are listed in each report.

        Engine
        vllm
        Protocol
        wbench-fixed-http-v1
        Dataset
        synthetic-tokenizer-v1:repeats=2
        Workload
        32768in-2048out
        ConcurrencyMean throughput (tok/s)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
        C162.4832.79 s3.6 s27.29433.31746bbc39 ↗

  2. clef-flash

    Cloudflare · 4 tested configurations · 1 GPU

    • vGPU-32GB · 32 GBbf16 · Text → Structured data · typed-invoices-v1 · Test duration 1 s
      Throughput · Batch 1
      10.3decisions/sec
      Best measured · Batch 4
      10.9decisions/sec
      Observed GPU occupancy
      18.2GiB
      P95 latency
      104 ms
      Use this recipe · Batch 4 ↗Evidence report and run availability ↗Same test on other GPUs
      Test details · 2 batch sizes

      Occupancy includes engine buffers and reserved KV cache. It is not the minimum VRAM required by the model.

      Reports combined counts reports, not inference requests. Request successes and failures are listed in each report.

      Engine
      clef-native
      Protocol
      native-forward-v1
      Dataset
      typed-invoices-v1
      Workload
      decision
      Batch sizeMean throughput (decisions/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
      B110.26104 ms—18.1767.911d1661ad ↗
      B410.89389 ms—18.4915611d1661ad ↗
    • 3 other test configurations on this GPU
      • bf16 · Text → Structured data · typed-invoices-v1 · Test duration 6 s
        Throughput · Batch 1
        10.6decisions/sec
        Best measured · Batch 4
        11decisions/sec
        Observed GPU occupancy
        18.2GiB
        P95 latency
        99 ms
        Use this recipe · Batch 4 ↗Evidence report and run availability ↗Same test on other GPUs
        Test details · 2 batch sizes

        Occupancy includes engine buffers and reserved KV cache. It is not the minimum VRAM required by the model.

        Reports combined counts reports, not inference requests. Request successes and failures are listed in each report.

        Engine
        clef-native · PyTorch 2.13.0+cu130
        Protocol
        native-forward-v1
        Dataset
        typed-invoices-v1
        Workload
        decision
        Batch sizeMean throughput (decisions/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
        B110.5699 ms—18.17206.1263f44036 ↗
        B411.04402 ms—18.55244.4263f44036 ↗
      • bf16 · Text → Structured data · typed-invoices-v1 · Test duration 6 s
        Throughput · Batch 1
        10.6decisions/sec
        Best measured · Batch 4
        11.1decisions/sec
        Observed GPU occupancy
        18.2GiB
        P95 latency
        103 ms
        Use this recipe · Batch 4 ↗Evidence report and run availability ↗Same test on other GPUs
        Test details · 2 batch sizes

        Occupancy includes engine buffers and reserved KV cache. It is not the minimum VRAM required by the model.

        Reports combined counts reports, not inference requests. Request successes and failures are listed in each report.

        Engine
        clef-native
        Protocol
        native-forward-v1
        Dataset
        typed-invoices-v1
        Workload
        decision
        Batch sizeMean throughput (decisions/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
        B110.56103 ms—18.17196.814eb73a4d ↗
        B411.1390 ms—18.56235.914eb73a4d ↗
      • bf16 · 1K → 1K tokens
        Throughput · Batch 1
        10.2decisions/sec
        Best measured · Batch 4
        10.9decisions/sec
        Observed GPU occupancy
        18.2GiB
        P95 latency
        104 ms
        Use this recipe · Batch 4 ↗Evidence report and run availability ↗Same test on other GPUs
        Test details · 2 batch sizes

        Occupancy includes engine buffers and reserved KV cache. It is not the minimum VRAM required by the model.

        Reports combined counts reports, not inference requests. Request successes and failures are listed in each report.

        Engine
        vLLM
        Protocol
        native-forward-v1
        Dataset
        typed-invoices-v1
        Workload
        decision
        Batch sizeMean throughput (decisions/sec)P95TTFTObserved occupancy (GiB)GPU power (W)Reports combinedView report
        B110.19104 ms—18.1769.81caef7154 ↗
        B410.88388 ms—18.49152.81caef7154 ↗
Models with reports but no qualified results · 72

Their original reports remain available on each model page. They did not qualify for comparison; each report states why.

Models with reports but no qualified results · 72 rows
ModelTest task
AutoGLM-Phone-9BZhipuAI/AutoGLM-Phone-9BText generation
DASD-4B-ThinkingAlibaba-Apsara/DASD-4B-ThinkingText generation
DeepSeek-OCR-2deepseek-ai/DeepSeek-OCR-2Text generation
DeepSeek-R1-Distill-Llama-8Bdeepseek-ai/DeepSeek-R1-Distill-Llama-8BText generation
DeepSeek-R1-Distill-Qwen-1.5Bdeepseek-ai/DeepSeek-R1-Distill-Qwen-1.5BText generation
DeepSeek-R1-Distill-Qwen-7Bdeepseek-ai/DeepSeek-R1-Distill-Qwen-7BText generation
dots.ocrrednote-hilab/dots.ocrText generation
Gemma-3-270mgoogle/gemma-3-270mText generation
Hunyuan-1.8B-FP8/root/.cache/modelscope/hub/models/Tencent-Hunyuan/HY-MT1___5-1___8B-FP8Text generation
Hunyuan-7B-FP8/root/.cache/modelscope/hub/models/Tencent-Hunyuan/HY-MT1___5-7B-FP8Text generation
HunyuanOCRTencent-Hunyuan/HunyuanOCRText generation
HY-MT1.5-1.8B-Q4_K_M/root/.cache/modelscope/hub/models/Tencent-Hunyuan/HY-MT1___5-1___8B-GGUF/HY-MT1.5-1.8B-Q4_K_M.ggufText generation
HY-MT1.5-1.8B-Q6_KTencent-Hunyuan/HY-MT1.5-1.8B-GGUF/HY-MT1.5-1.8B-Q6_K.ggufText generation
HY-MT1.5-1.8B-Q8_0Tencent-Hunyuan/HY-MT1.5-1.8B-GGUF/HY-MT1.5-1.8B-Q8_0.ggufText generation
HY-MT1.5-7B-Q4_K_M/root/.cache/modelscope/hub/models/Tencent-Hunyuan/HY-MT1___5-7B-GGUF/HY-MT1.5-7B-Q4_K_M.ggufText generation
HY-MT1.5-7B-Q6_KTencent-Hunyuan/HY-MT1.5-7B-GGUF/HY-MT1.5-7B-Q6_K.ggufText generation
HY-MT1.5-7B-Q8_0Tencent-Hunyuan/HY-MT1.5-7B-GGUF/HY-MT1.5-7B-Q8_0.ggufText generation
LightOnOCR-2-1Blightonai/LightOnOCR-2-1BText generation
MAI-UI-8BTongyi-MAI/MAI-UI-8BText generation
MiniCPM-o-4_5openbmb/MiniCPM-o-4_5Text generation
Nanbeige4-3B-Thinking-2511Nanbeige/Nanbeige4-3B-Thinking-2511Text generation
PaddleOCR-VL-1.5PaddlePaddle/PaddleOCR-VL-1.5Text generation
Qwen2.5-0.5B-Instruct-AWQ/root/.cache/modelscope/hub/models/Qwen/Qwen2___5-0___5B-Instruct-AWQText generation
qwen2.5-0.5b-instruct-q4_k_m.ggufQwen/Qwen2.5-0.5B-Instruct-GGUF/qwen2.5-0.5b-instruct-q4_k_m.ggufText generation
Qwen2.5-1.5B-Instruct-AWQ/root/.cache/modelscope/hub/models/Qwen/Qwen2___5-1___5B-Instruct-AWQText generation
Qwen2.5-7B-Instruct-AWQ/root/.cache/modelscope/hub/models/Qwen/Qwen2___5-7B-Instruct-AWQText generation
Qwen2.5-72B-AWQQwen/Qwen2.5-72B-Instruct-AWQText generation
Qwen3-0.6BQwen/Qwen3-0.6BText generation
Qwen3-0.6B-FP8Qwen/Qwen3-0.6B-FP8Text generation
Qwen3-0.6B-Q8Qwen/Qwen3-0.6B-GGUF/Qwen3-0.6B-Q8_0.ggufText generation
Qwen3-1.7BQwen/Qwen3-1.7BText generation
Qwen3-1.7B-FP8Qwen/Qwen3-1.7B-FP8Text generation
Qwen3-1.7B-Q8Qwen/Qwen3-1.7B-GGUF/Qwen3-1.7B-Q8_0.ggufText generation
Qwen3-4BQwen/Qwen3-4BText generation
Qwen3-4B-AWQQwen/Qwen3-4B-AWQText generation
Qwen3-4B-FP8Qwen/Qwen3-4B-FP8Text generation
Qwen3-4B-Instruct-2507-FP8Qwen/Qwen3-4B-Instruct-2507-FP8Text generation
Qwen3-4B-Q4Qwen/Qwen3-4B-GGUF/Qwen3-4B-Q4_K_M.ggufText generation
Qwen3-4B-Q5_K_MQwen/Qwen3-4B-GGUF/Qwen3-4B-Q5_K_M.ggufText generation
Qwen3-4B-Q8_0Qwen/Qwen3-4B-GGUF/Qwen3-4B-Q8_0.ggufText generation
Qwen3-4B-Thinking-2507-FP8Qwen/Qwen3-4B-Thinking-2507-FP8Text generation
Qwen3-8B-AWQQwen/Qwen3-8B-AWQText generation
Qwen3-8B-FP8Qwen/Qwen3-8B-FP8Text generation
Qwen3-8B-Q4Qwen/Qwen3-8B-GGUF/Qwen3-8B-Q4_K_M.ggufText generation
Qwen3-8B-Q5_K_MQwen/Qwen3-8B-GGUF/Qwen3-8B-Q5_K_M.ggufText generation
Qwen3-8B-Q8_0Qwen/Qwen3-8B-GGUF/Qwen3-8B-Q8_0.ggufText generation
Qwen3-14B-AWQQwen/Qwen3-14B-AWQText generation
Qwen3-14B-FP8Qwen/Qwen3-14B-FP8Text generation
Qwen3-14B-Q4Qwen/Qwen3-14B-GGUF/Qwen3-14B-Q4_K_M.ggufText generation
Qwen3-14B-Q5_K_MQwen/Qwen3-14B-GGUF/Qwen3-14B-Q5_K_M.ggufText generation
Qwen3-14B-Q8_0Qwen/Qwen3-14B-GGUF/Qwen3-14B-Q8_0.ggufText generation
Qwen3-30B-A3B-FP8Qwen/Qwen3-30B-A3B-Instruct-2507-FP8Text generation
Qwen3-30B-A3B-FP8/root/.cache/modelscope/hub/models/Qwen/Qwen3-30B-A3B-FP8Text generation
Qwen3-30B-A3B-Q4_K_MQwen/Qwen3-30B-A3B-GGUF/Qwen3-30B-A3B-Q4_K_M.ggufText generation
Qwen3-30B-A3B-Q8_0Qwen/Qwen3-30B-A3B-GGUF/Qwen3-30B-A3B-Q8_0.ggufText generation
Qwen3-30B-A3B-Thinking-FP8Qwen/Qwen3-30B-A3B-Thinking-2507-FP8Text generation
Qwen3-32B-AWQQwen/Qwen3-32B-AWQText generation
Qwen3-32B-FP8Qwen/Qwen3-32B-FP8Text generation
Qwen3-32B-Q4Qwen/Qwen3-32B-GGUF/Qwen3-32B-Q4_K_M.ggufText generation
Qwen3-32B-Q5_K_MQwen/Qwen3-32B-GGUF/Qwen3-32B-Q5_K_M.ggufText generation
Qwen3-VL-2B-Instruct-FP8Qwen/Qwen3-VL-2B-Instruct-FP8Text generation
Qwen3-VL-4B-Instruct-FP8Qwen/Qwen3-VL-4B-Instruct-FP8Text generation
Qwen3-VL-8B-Instruct-FP8Qwen/Qwen3-VL-8B-Instruct-FP8Text generation
Qwen3-VL-30B-A3B-FP8Qwen/Qwen3-VL-30B-A3B-Instruct-FP8Text generation
Qwen3-VL-32B-FP8Qwen/Qwen3-VL-32B-Instruct-FP8Text generation
Qwen3.5-0.8BQwen/Qwen3.5-0.8BText generation
Qwen3.5-2BQwen/Qwen3.5-2BText generation
Qwen3.5-4BQwen/Qwen3.5-4BText generation
Qwen3.5-9BQwen/Qwen3.5-9BText generation
Qwen3.5-27B-FP8Qwen/Qwen3.5-27B-FP8Text generation
Qwen3.8-27B-NVFP4RadixArk/Qwen3.8-27B-NVFP4Text generation
Step3-VL-10Bstepfun-ai/Step3-VL-10BText generation