Skip to content

Open models to fine-tune and serve

Model families

28 families

  • Qwen3

    Qwen

    Supported
    • Dense
    • Text
    Sizes
    • 0.6B
    • 1.7B
    • 4B
    • 8B
    • 14B
    • 32B
    Suggested sizes
    LoRA from 1 GPU, about 9 GB
    Serves
    Adapter, base model

    Qwen3

    Qwen

    Supported
    • Qwen/Qwen3-0.6B

      0.6BSupported

      LoRA
      1 GPU, about 6 GB
      Serve
      1, 2, 4 or 8 GPUs, about 4 GB in all
    • Qwen/Qwen3-1.7B

      1.7BSupported

      LoRA
      1 GPU, about 9 GB
      Serve
      1, 2, 4 or 8 GPUs, about 7 GB in all
    • Qwen/Qwen3-4B-Instruct-2507

      4BSupported

      LoRA
      1 GPU, about 14 GB
      Serve
      1, 2, 4 or 8 GPUs, about 12 GB in all
    • Qwen/Qwen3-4B

      4BSupported

      LoRA
      1 GPU, about 14 GB
      Serve
      1, 2, 4 or 8 GPUs, about 12 GB in all
    • Qwen/Qwen3-8B

      8BSupported

      LoRA
      1 GPU, about 24 GB
      Serve
      1, 2, 4 or 8 GPUs, about 22 GB in all
    • Qwen/Qwen3-14B

      14BSupported

      LoRA
      1 GPU, about 40 GB
      Serve
      1, 2, 4 or 8 GPUs, about 38 GB in all
    • Qwen/Qwen3-32B

      32BSupported

      LoRA
      1 GPU, about 83 GB
      Serve
      1, 2, 4 or 8 GPUs, about 81 GB in all
    Fine-tune Qwen3
  • Qwen3 MoE

    Qwen

    Supported
    • Mixture of experts
    • Text
    Sizes
    • 30B, 3B active
    Suggested sizes
    LoRA from 1 GPU, about 78 GB
    Serves
    Adapter, base model

    Qwen3 MoE

    Qwen

    Supported
    • Qwen/Qwen3-30B-A3B-Instruct-2507

      30B, 3B activeSupported

      LoRA
      1 GPU, about 78 GB
      Serve
      1, 2, 4 or 8 GPUs, about 76 GB in all
    • Qwen/Qwen3-30B-A3B-Thinking-2507

      30B, 3B activeSupported

      LoRA
      1 GPU, about 78 GB
      Serve
      1, 2, 4 or 8 GPUs, about 76 GB in all
    • Qwen/Qwen3-Coder-30B-A3B-Instruct

      30B, 3B activeSupported

      LoRA
      1 GPU, about 78 GB
      Serve
      1, 2, 4 or 8 GPUs, about 76 GB in all
    Fine-tune Qwen3 MoE
  • Qwen3.5, Qwen3.6 and Qwen3.8

    Qwen

    Supported
    • Dense
    • Text and images
    Sizes
    • 9B
    • 27B
    Suggested sizes
    LoRA from 1 GPU, about 71 GB
    Serves
    Adapter, base model

    Qwen3.5, Qwen3.6 and Qwen3.8

    Qwen

    Supported
    • Qwen/Qwen3.5-9B

      9BSupported

      LoRA
      1 GPU, about 27 GB
      Serve
      1, 2, 4 or 8 GPUs, about 25 GB in all
    • Qwen/Qwen3.8-27B

      27BSupported

      LoRA
      1 GPU, about 71 GB
      Serve
      1, 2, 4 or 8 GPUs, about 69 GB in all
    • Qwen/Qwen3.6-27B

      27BSupported

      LoRA
      1 GPU, about 71 GB
      Serve
      1, 2, 4 or 8 GPUs, about 69 GB in all
    Fine-tune Qwen3.5, Qwen3.6 and Qwen3.8
  • Qwen3.5 and Qwen3.6 MoE

    Qwen

    Supported
    • Mixture of experts
    • Text and images
    Sizes
    • 35B, 3B active
    • 122B, 10B active, serving only
    Suggested sizes
    LoRA from 1 GPU, about 91 GB
    Serves
    Adapter, base model

    Qwen3.5 and Qwen3.6 MoE

    Qwen

    Supported
    • Qwen/Qwen3.6-35B-A3B

      35B, 3B activeSupported

      LoRA
      1 GPU, about 91 GB
      Serve
      1, 2, 4 or 8 GPUs, about 89 GB in all
    • Qwen/Qwen3.5-35B-A3B

      35B, 3B activeSupported

      LoRA
      1 GPU, about 91 GB
      Serve
      1, 2, 4 or 8 GPUs, about 89 GB in all
    • Qwen/Qwen3.5-122B-A10B

      122B, 10B activeSupportedServing only

      LoRA
      -
      Serve
      2, 4 or 8 GPUs, about 305 GB in all
    Fine-tune Qwen3.5 and Qwen3.6 MoE
  • Qwen3.8 Flash-Next

    Qwen

    Experimental
    • Mixture of experts
    • Text and images
    Sizes
    • 180B
    Suggested sizes
    LoRA from 8 x NVIDIA B200
    Serves
    Adapter, base model

    Qwen3.8 Flash-Next

    Qwen

    Experimental
    • Qwen/Qwen3.8-Flash-Next

      180BExperimental

      LoRA
      8 x NVIDIA B200
      Serve
      4 or 8 GPUs, about 438 GB in all
    Fine-tune Qwen3.8 Flash-Next
  • gpt-oss

    OpenAI

    Supported
    • Mixture of experts
    • Text
    Sizes
    • 20B
    • 120B, serving only
    Suggested sizes
    LoRA from 1 GPU, about 55 GB
    Serves
    Adapter, base model

    gpt-oss

    OpenAI

    Supported
    • unsloth/gpt-oss-20b-BF16

      20BSupported

      LoRA
      1 GPU, about 55 GB
      Serve
      1, 2, 4 or 8 GPUs, about 53 GB in all
    • openai/gpt-oss-20b

      20BSupportedServing only

      LoRA
      -
      Serve
      1, 2, 4 or 8 GPUs, about 17 GB in all
    • unsloth/gpt-oss-120b-BF16

      120BSupportedServing only

      LoRA
      -
      Serve
      2, 4 or 8 GPUs, about 285 GB in all
    Fine-tune gpt-oss
  • GLM-4.7 Flash

    Z.ai

    Supported
    • Mixture of experts
    • Text
    Sizes
    • 31B
    Suggested sizes
    LoRA from 1 GPU, about 80 GB
    Serves
    Adapter, base model

    GLM-4.7 Flash

    Z.ai

    Supported
    • zai-org/GLM-4.7-Flash

      31BSupported

      LoRA
      1 GPU, about 80 GB
      Serve
      1, 2 or 4 GPUs, about 78 GB in all
    Fine-tune GLM-4.7 Flash
  • Gemma 4

    Google

    Supported
    • Dense and mixture of experts
    • Text and images
    Sizes
    • E2B
    • E4B
    • 12B
    • 26B, 4B active
    • 31B
    Suggested sizes
    LoRA from 1 GPU, about 33 GB
    Serves
    Adapter, base model

    Gemma 4

    Google

    Supported
    • google/gemma-4-E2B-it

      E2BSupported

      LoRA
      1 GPU, about 16 GB
      Serve
      1, 2 or 4 GPUs, about 14 GB in all
    • google/gemma-4-E4B-it

      E4BSupported

      LoRA
      1 GPU, about 23 GB
      Serve
      1, 2 or 4 GPUs, about 21 GB in all
    • google/gemma-4-12B-it

      12BSupported

      LoRA
      1 GPU, about 33 GB
      Serve
      1, 2, 4 or 8 GPUs, about 31 GB in all
    • google/gemma-4-26B-A4B-it

      26B, 4B activeSupported

      LoRA
      1 GPU, about 67 GB
      Serve
      1, 2, 4 or 8 GPUs, about 65 GB in all
    • google/gemma-4-31B-it

      31BSupported

      LoRA
      1 GPU, about 80 GB
      Serve
      1, 2, 4 or 8 GPUs, about 78 GB in all
    Fine-tune Gemma 4
  • Laguna 2.1

    Poolside

    Supported
    • Mixture of experts
    • Text
    Sizes
    • 33B
    • 118B, serving only
    Suggested sizes
    LoRA from 1 GPU, about 85 GB
    Serves
    Adapter, base model

    Laguna 2.1

    Poolside

    Supported
    • poolside/Laguna-XS-2.1

      33BSupported

      LoRA
      1 GPU, about 85 GB
      Serve
      1, 2, 4 or 8 GPUs, about 83 GB in all
    • poolside/Laguna-S-2.1

      118BSupportedServing only

      LoRA
      -
      Serve
      2, 4 or 8 GPUs, about 287 GB in all
    Fine-tune Laguna 2.1
  • LFM2 MoE

    Liquid AI

    Supported
    • Mixture of experts
    • Text
    Sizes
    • 8B, 1B active
    • 24B, 2B active
    Suggested sizes
    LoRA from 1 GPU, about 25 GB
    Serves
    Adapter, base model

    LFM2 MoE

    Liquid AI

    Supported
    • LiquidAI/LFM2.5-8B-A1B

      8B, 1B activeSupported

      LoRA
      1 GPU, about 25 GB
      Serve
      1, 2, 4 or 8 GPUs, about 23 GB in all
    • LiquidAI/LFM2-24B-A2B

      24B, 2B activeSupported

      LoRA
      1 GPU, about 62 GB
      Serve
      1, 2, 4 or 8 GPUs, about 60 GB in all
    Fine-tune LFM2 MoE
  • NVIDIA Nemotron 3.5

    NVIDIA

    Supported
    • Mixture of experts
    • Text
    Sizes
    • 30B, 3B active
    • 120B, 12B active, serving only
    Suggested sizes
    LoRA from 1 GPU, about 81 GB
    Serves
    Adapter, base model

    NVIDIA Nemotron 3.5

    NVIDIA

    Supported
    • nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16

      30B, 3B activeSupported

      LoRA
      1 GPU, about 81 GB
      Serve
      1, 2, 4 or 8 GPUs, about 79 GB in all
    • nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16

      120B, 12B activeSupportedServing only

      LoRA
      -
      Serve
      2, 4 or 8 GPUs, about 303 GB in all
    Fine-tune NVIDIA Nemotron 3.5
  • Llama

    Meta

    Supported
    • Dense
    • Text
    Sizes
    • 1B
    • 3B
    • 8B
    Suggested sizes
    LoRA from 1 GPU, about 7 GB
    Serves
    Adapter, base model

    Llama

    Meta

    Supported
    • unsloth/Llama-3.2-1B-Instruct

      1BSupported

      LoRA
      1 GPU, about 7 GB
      Serve
      1, 2, 4 or 8 GPUs, about 5 GB in all
    • unsloth/Llama-3.2-3B-Instruct

      3BSupported

      LoRA
      1 GPU, about 12 GB
      Serve
      1, 2, 4 or 8 GPUs, about 10 GB in all
    • unsloth/Llama-3.1-8B-Instruct

      8BSupported

      LoRA
      1 GPU, about 23 GB
      Serve
      1, 2, 4 or 8 GPUs, about 21 GB in all
    Fine-tune Llama
  • SmolLM2

    Hugging Face

    Supported
    • Dense
    • Text
    Sizes
    • 360M
    • 1.7B
    Suggested sizes
    LoRA from 1 GPU, about 8 GB
    Serves
    Adapter, base model

    SmolLM2

    Hugging Face

    Supported
    • HuggingFaceTB/SmolLM2-360M-Instruct

      360MSupported

      LoRA
      1 GPU, about 5 GB
      Serve
      1 GPU, about 3 GB in all
    • HuggingFaceTB/SmolLM2-1.7B-Instruct

      1.7BSupported

      LoRA
      1 GPU, about 8 GB
      Serve
      1, 2, 4 or 8 GPUs, about 6 GB in all
    Fine-tune SmolLM2
  • Mistral

    Mistral AI

    Supported
    • Dense
    • Text
    Sizes
    • 7B
    Suggested sizes
    LoRA from 1 GPU, about 22 GB
    Serves
    Adapter, base model

    Mistral

    Mistral AI

    Supported
    • mistralai/Mistral-7B-Instruct-v0.3

      7BSupported

      LoRA
      1 GPU, about 22 GB
      Serve
      1, 2, 4 or 8 GPUs, about 20 GB in all
    Fine-tune Mistral
  • Phi-3

    Microsoft

    Supported
    • Dense
    • Text
    Sizes
    • 3.8B
    Suggested sizes
    LoRA from 1 GPU, about 13 GB
    Serves
    Adapter, base model

    Phi-3

    Microsoft

    Supported
    • microsoft/Phi-3-mini-4k-instruct

      3.8BSupported

      LoRA
      1 GPU, about 13 GB
      Serve
      1, 2, 4 or 8 GPUs, about 11 GB in all
    Fine-tune Phi-3
  • DeepSeek R1 distills

    DeepSeek

    Supported
    • Dense
    • Text
    Sizes
    • 1.5B
    • 7B
    • 8B
    • 14B
    • 32B
    • 70B
    Suggested sizes
    LoRA from 1 GPU, about 8 GB
    Serves
    Adapter, base model

    DeepSeek R1 distills

    DeepSeek

    Supported
    • deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

      1.5BSupported

      LoRA
      1 GPU, about 8 GB
      Serve
      1, 2 or 4 GPUs, about 6 GB in all
    • deepseek-ai/DeepSeek-R1-Distill-Qwen-7B

      7BSupported

      LoRA
      1 GPU, about 22 GB
      Serve
      1, 2 or 4 GPUs, about 20 GB in all
    • deepseek-ai/DeepSeek-R1-Distill-Llama-8B

      8BSupported

      LoRA
      1 GPU, about 23 GB
      Serve
      1, 2, 4 or 8 GPUs, about 21 GB in all
    • deepseek-ai/DeepSeek-R1-0528-Qwen3-8B

      8BSupported

      LoRA
      1 GPU, about 24 GB
      Serve
      1, 2, 4 or 8 GPUs, about 22 GB in all
    • deepseek-ai/DeepSeek-R1-Distill-Qwen-14B

      14BSupported

      LoRA
      1 GPU, about 40 GB
      Serve
      1, 2, 4 or 8 GPUs, about 38 GB in all
    • deepseek-ai/DeepSeek-R1-Distill-Qwen-32B

      32BSupported

      LoRA
      1 GPU, about 83 GB
      Serve
      1, 2, 4 or 8 GPUs, about 81 GB in all
    • deepseek-ai/DeepSeek-R1-Distill-Llama-70B

      70BSupported

      LoRA
      1 GPU, about 175 GB
      Serve
      1, 2, 4 or 8 GPUs, about 173 GB in all
    Fine-tune DeepSeek R1 distills
  • DeepSeek V2 Lite

    DeepSeek

    Supported
    • Mixture of experts
    • Text
    Sizes
    • 16B
    Suggested sizes
    LoRA from 1 GPU, about 42 GB
    Serves
    Adapter, base model

    DeepSeek V2 Lite

    DeepSeek

    Supported
    • deepseek-ai/DeepSeek-V2-Lite-Chat

      16BSupported

      LoRA
      1 GPU, about 42 GB
      Serve
      1, 2, 4 or 8 GPUs, about 40 GB in all
    • deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct

      16BSupported

      LoRA
      1 GPU, about 42 GB
      Serve
      1, 2, 4 or 8 GPUs, about 40 GB in all
    Fine-tune DeepSeek V2 Lite
  • Qwen2.5-VL

    Qwen

    Supported
    • Dense
    • Text and images
    Sizes
    • 3B
    • 7B
    Suggested sizes
    LoRA from 1 GPU, about 13 GB
    Serves
    Adapter, base model

    Qwen2.5-VL

    Qwen

    Supported
    • Qwen/Qwen2.5-VL-3B-Instruct

      3BSupported

      LoRA
      1 GPU, about 13 GB
      Serve
      1, 2, 4 or 8 GPUs, about 11 GB in all
    • Qwen/Qwen2.5-VL-7B-Instruct

      7BSupported

      LoRA
      1 GPU, about 24 GB
      Serve
      1, 2 or 4 GPUs, about 22 GB in all
    Fine-tune Qwen2.5-VL
  • Qwen3-VL

    Qwen

    Supported
    • Dense
    • Text and images
    Sizes
    • 4B
    • 8B
    Suggested sizes
    LoRA from 1 GPU, about 15 GB
    Serves
    Adapter, base model

    Qwen3-VL

    Qwen

    Supported
    • Qwen/Qwen3-VL-4B-Instruct

      4BSupported

      LoRA
      1 GPU, about 15 GB
      Serve
      1, 2, 4 or 8 GPUs, about 13 GB in all
    • Qwen/Qwen3-VL-8B-Instruct

      8BSupported

      LoRA
      1 GPU, about 25 GB
      Serve
      1, 2, 4 or 8 GPUs, about 23 GB in all
    Fine-tune Qwen3-VL
  • Gemma 3

    Google

    Supported
    • Dense
    • Text and images
    Sizes
    • 4B
    • 12B
    • 27B
    Suggested sizes
    LoRA from 1 GPU, about 14 GB
    Serves
    Adapter, base model

    Gemma 3

    Google

    Supported
    • unsloth/gemma-3-4b-it

      4BSupported

      LoRA
      1 GPU, about 14 GB
      Serve
      1, 2, 4 or 8 GPUs, about 12 GB in all
    • unsloth/gemma-3-12b-it

      12BSupported

      LoRA
      1 GPU, about 34 GB
      Serve
      1, 2, 4 or 8 GPUs, about 32 GB in all
    • unsloth/gemma-3-27b-it

      27BSupported

      LoRA
      1 GPU, about 70 GB
      Serve
      1, 2, 4 or 8 GPUs, about 68 GB in all
    Fine-tune Gemma 3
  • SmolVLM

    Hugging Face

    Supported
    • Dense
    • Text and images
    Sizes
    • 2B
    Suggested sizes
    LoRA from 1 GPU, about 9 GB
    Serves
    Adapter, base model

    SmolVLM

    Hugging Face

    Supported
    • HuggingFaceTB/SmolVLM-Instruct

      2BSupported

      LoRA
      1 GPU, about 9 GB
      Serve
      1, 2, 4 or 8 GPUs, about 7 GB in all
    Fine-tune SmolVLM
  • ModernBERT and Ettin (decision models)

    Answer.AI and Johns Hopkins University

    Supported
    • Dense
    • Text
    • Decision model
    Sizes
    • 150M
    • 400M
    • 396M
    Suggested sizes
    LoRA from 1 GPU, about 7 GB
    Serves
    Adapter, base model

    ModernBERT and Ettin (decision models)

    Answer.AI and Johns Hopkins University

    Supported
    • answerdotai/ModernBERT-base

      150MSupported

      LoRA
      1 GPU, about 7 GB
      Serve
      -
    • jhu-clsp/ettin-encoder-400m

      400MSupported

      LoRA
      1 GPU, about 8 GB
      Serve
      -
    • skundu42/kev

      396MSupported

      LoRA
      1 GPU, about 8 GB
      Serve
      1 GPU, about 4 GB in all
    Fine-tune ModernBERT and Ettin
  • Mistral Small 4

    Mistral AI

    Supported
    • Mixture of experts
    • Text
    • Serving only
    Sizes
    • 119B, serving only
    LoRA from
    -
    Serves
    Base model

    Mistral Small 4

    Mistral AI

    Supported
    • mistralai/Mistral-Small-4-119B-2603

      119BSupportedServing only

      LoRA
      -
      Serve
      1, 2, 4 or 8 GPUs, about 132 GB in all
  • Command A+

    Cohere

    Supported
    • Mixture of experts
    • Text
    • Serving only
    Sizes
    • 219B, serving only
    LoRA from
    -
    Serves
    Base model

    Command A+

    Cohere

    Supported
    • CohereLabs/command-a-plus-05-2026-bf16

      219BSupportedServing only

      LoRA
      -
      Serve
      4 or 8 GPUs, about 532 GB in all
  • Inkling

    Thinking Machines

    Supported
    • Mixture of experts
    • Text
    • Serving only
    Sizes
    • 266B, serving only
    LoRA from
    -
    Serves
    Base model

    Inkling

    Thinking Machines

    Supported
    • thinkingmachines/Inkling-Small

      266BSupportedServing only

      LoRA
      -
      Serve
      4 or 8 GPUs, about 646 GB in all
  • GLM-5.3 Flash

    Z.ai

    Supported
    • Mixture of experts
    • Text
    • Serving only
    Sizes
    • 321B, serving only
    LoRA from
    -
    Serves
    Base model

    GLM-5.3 Flash

    Z.ai

    Supported
    • zai-org/GLM-5.3-Flash

      321BSupportedServing only

      LoRA
      -
      Serve
      2, 4 or 8 GPUs, about 354 GB in all
  • Step 3.7 Flash

    StepFun

    Supported
    • Mixture of experts
    • Text
    • Serving only
    Sizes
    • 201B, serving only
    LoRA from
    -
    Serves
    Base model

    Step 3.7 Flash

    StepFun

    Supported
    • stepfun-ai/Step-3.7-Flash

      201BSupportedServing only

      LoRA
      -
      Serve
      4 or 8 GPUs, about 490 GB in all
  • Qwen3-Next

    Qwen

    Supported
    • Mixture of experts
    • Text
    • Serving only
    Sizes
    • 80B, 3B active, serving only
    LoRA from
    -
    Serves
    Base model

    Qwen3-Next

    Qwen

    Supported
    • Qwen/Qwen3-Next-80B-A3B-Instruct

      80B, 3B activeSupportedServing only

      LoRA
      -
      Serve
      2, 4 or 8 GPUs, about 199 GB in all

Model check

What about other models?

Any other Hugging Face chat model with a chat template is checked when you create a project. A model outside these families passes with a warning: it may not train or serve. See the supported-models docs.

Start with one dataset.

Create an account and import your data, or book a demo and we will show you the product.