Skip to main content

Local AI · historical measurements (legacy, read-only)

Find a local AI model for your GPU

Pick a task, your GPU, and a context length. Speeds are earlier Bench-Agent measurements on that exact GPU, model file, and context — not a UOS index and not used in upgrade decisions. Model quality and license are separate questions.

A model's published context limit does not prove it fits your GPU at that length.

Measured on real hardware

Choose your GPU above to see models measured on it. Measurements are never assigned to a different or unknown GPU.

Model candidates for coding at 32K

A shortlist, not a ranking. Model quality, GPU fit, context, local speed and license are separate — there is no single AI score.

Model candidates for coding at 32K
ModelPrimary useQuality evidenceGPU fitContextLocal speedLicenseDetails
DeepSeek R1 Distill · 14BReasoning, Math2 reportedNot verifiedUp to 128K documentedNot measuredNeeds reviewDetails for DeepSeek R1 Distill
Gemma 4 · 11.95BCoding, Reasoning1 reportedNot verifiedUp to 256K documentedNot measuredNeeds reviewDetails for Gemma 4
Mistral Small 3.2 · 24BCoding, Reasoning2 reportedNot verifiedUp to 128K documentedNot measuredNeeds reviewDetails for Mistral Small 3.2
Llama 3.1 · 8BCoding, Reasoning2 reportedNot verifiedUp to 128K documentedNot measuredNeeds reviewDetails for Llama 3.1
  • DeepSeek R1 Distill · 14B

    Reasoning, Math

    Quality
    2 reported
    GPU fit
    Not verified
    Local speed
    Not measured
    License
    Needs review
    Details
  • Gemma 4 · 11.95B

    Coding, Reasoning

    Quality
    1 reported
    GPU fit
    Not verified
    Local speed
    Not measured
    License
    Needs review
    Details
  • Mistral Small 3.2 · 24B

    Coding, Reasoning

    Quality
    2 reported
    GPU fit
    Not verified
    Local speed
    Not measured
    License
    Needs review
    Details
  • Llama 3.1 · 8B

    Coding, Reasoning

    Quality
    2 reported
    GPU fit
    Not verified
    Local speed
    Not measured
    License
    Needs review
    Details

Model details

DeepSeek R1 Distill · 14B

Category-specific provider quality facts exist, but this GPU/context has no reviewed runtime fit evidence.

Context: up to 128K documented, not verified on your GPU.

  • Coding · LiveCodeBench: 53.1 Provider reported
  • Coding · Codeforces rating: 1481 Provider reported

Backend: not documented · Context note: Base model config publishes max_position_embeddings 131,072; practical memory fit by context is unknown. · Base license: MIT (provider-reported); file license: not documented.

Official model source
Gemma 4 · 11.95B

Category-specific provider quality facts exist, but this GPU/context has no reviewed runtime fit evidence.

Context: up to 256K documented, not verified on your GPU.

  • Coding · LiveCodeBench v6 · Gemma 4 12B Unified: 72 Provider reported

Backend: not documented · Context note: Google model card reports a 256K context window. This is a documented model limit, not proof of local memory fit. · Base license: Apache-2.0 (Google repository-declared); file license: not documented.

Official model source
Mistral Small 3.2 · 24B

Category-specific provider quality facts exist, but this GPU/context has no reviewed runtime fit evidence.

Context: up to 128K documented, not verified on your GPU.

  • Coding · HumanEval Plus · Pass@5 · Small 3.2 24B Instruct: 92.9 Provider reported
  • Coding · MBPP Plus · Pass@5 · Small 3.2 24B Instruct: 78.33 Provider reported

Backend: not documented · Context note: Official model card/config describe 131,072 tokens. The card states approximately 55 GB GPU memory for bf16/fp16; quantized local fit is not established here. · Base license: Apache-2.0 (official model repository-declared); file license: not documented.

Official model source
Llama 3.1 · 8B

Category-specific provider quality facts exist, but this GPU/context has no reviewed runtime fit evidence.

Context: up to 128K documented, not verified on your GPU.

  • Coding · HumanEval · zero-shot pass@1 · Llama 3.1 8B Instruct: 72.6 Provider reported
  • Coding · MBPP ++ base version · zero-shot pass@1 · Llama 3.1 8B Instruct: 72.8 Provider reported

Backend: not documented · Context note: Meta model card reports 128K context. Its gated repository requires access approval; the custom Llama 3.1 Community License is retained as review-required for commercial-use filtering. No quant artifact or local fit is asserted. · Base license: Llama 3.1 Community License (custom; terms apply); file license: not documented.

Official model source

Measured fit can narrow a shortlist. Model quality and your own preferences still decide which model is best for you.