Menu
mutinaiPreview: Public preview · figures are illustrative fixture data, attributed to their source · data to May 2025 · read-only
What can I run?
Open models · sources checked 24 Sept 2026

AI you can download, run and keep.

Open models are published as files anyone can download: you run them on your own hardware instead of sending your work to someone else’s. Mutinai tracks the models, the machines they fit and the tools that run them, and labels every figure with where it came from.

What can I run?What’s happening right now
  • 25open-weight models
  • 8releases & launches in 30 days
  • Best open GPQA Diamond 71.5%(DeepSeek-R1)
  • 14tools & runtimes

What’s happening right now

Full timeline →
Top story
runtime release ·

llama.cpp v0.5.0

This release focuses on backend performance and correctness, broader model coverage, and more robust server/router operation. It adds HRM-Text (DFM Mimir 1B) support, MiMo-V2.6 and HunyuanOCR conversion support, ggml 0.25.0 backend improvements, multi-address HTTP binding, image

Tool llama.cpp

Latest

  1. Event
    Accelerating vision-language models with LFM2.5-VL-DSpark
    announcement
  2. Event
    How to Use NVIDIA Warp and MjWarp to Accelerate Robotics Simulation and Learning Workflows
    announcement
  3. Event
    **Know Who Spoke When: Build Real-Time, Multi-Speaker AI with NVIDIA Nemotron 3 Diarization**
    announcement
  4. Event
    Pruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem
    announcement
  5. Toolruntime release
  6. Eventannouncement
  7. Toolruntime release
  8. Event
    Your Agent Aced the Task. Will It Do It Again?
    announcement

Models worth knowing

Open models arrive in families. Start with the family and who builds it, then open a release to see its sizes and variants.

All models →

  • Qwen

    Qwen Team (Alibaba Cloud)
    Known forCoding, reasoning and tool use
    596M – 235B parametersHow many numbers the model learned during training — the usual rough measure of its size.Dense and mixture of expertsA model split into parts where only a few run for each word, so it answers faster than its size suggests.Newest release 5 Aug 2026
    Qwen3.8 27BQwen3 0.6BQwen3 1.7Ball 12 →

  • Gemma

    Google DeepMind
    Known forImages, long documents and many languages
    9.24B – 27.4B parametersHow many numbers the model learned during training — the usual rough measure of its size.DenseNewest release 12 Mar 2025
    Gemma 3 27BGemma 2 9BGemma 2 27B

  • Mistral

    Mistral AI
    Known forTool use and many languages
    7.25B – 23.6B parametersHow many numbers the model learned during training — the usual rough measure of its size.DenseNewest release 30 Jan 2025
    Mistral Small 24BMistral 7B

  • DeepSeek-R1

    DeepSeek
    Known forCoding, reasoning and long documents
    671B parametersHow many numbers the model learned during training — the usual rough measure of its size.Mixture of expertsA model split into parts where only a few run for each word, so it answers faster than its size suggests.Newest release 20 Jan 2025
    DeepSeek-R1 671B

  • Phi

    Microsoft
    Known forCoding and reasoning
    14.7B parametersHow many numbers the model learned during training — the usual rough measure of its size.DenseNewest release 12 Dec 2024
    Phi-4 14B

  • Llama

    Meta
    Known forCoding, tool use and long documents
    8.03B – 70.6B parametersHow many numbers the model learned during training — the usual rough measure of its size.DenseNewest release 6 Dec 2024
    Llama 3.3 70BLlama 3.1 8BLlama 3.1 70B

How they compare

A benchmarkA fixed set of questions every model is given, so their scores can be compared on the same task. is a fixed set of questions every model answers, so scores can be compared. Positions below are relative to the best open result Mutinai holds for that test.

All benchmarks →

Coding best scores on HumanEval

  1. Qwen2.5-Coder 32B92.7
  2. Llama 3.3 70B88.4
  3. Qwen2.5 32B88.4

developer-reported what it measures →

Reasoning best scores on GPQA Diamond

  1. DeepSeek-R1 671B71.5
  2. Qwen3 30B-A3B65.8
  3. Qwen2.5 32B62.1

developer-reported what it measures →

Knowledge best scores on MMLU-Pro

  1. DeepSeek-R1 671B84.0
  2. Phi-4 14B70.4
  3. Qwen2.5 32B69.0

developer-reported what it measures →

Instruction following best scores on IFEval

  1. Llama 3.3 70B92.1
  2. Gemma 3 27B90.4
  3. Llama 3.1 8B80.4

developer-reported what it measures →

Runs on your own machine

  1. Qwen2.5-Coder 32B~21 GB to runruns well on 11 of 19 reference systems
  2. Qwen2.5-Math 7B~6 GB to runruns well on 17 of 19 reference systems
  3. Qwen3 30B-A3B~18 GB to runruns well on 11 of 19 reference systems

estimated smallest download at 8K contextHow much text the model can consider at once, counted in tokens — roughly ¾ of a word each.

Images and multimodal

1 tracked model accepts images, but Mutinai holds no multimodal benchmark results yet, so there is nothing to rank here. See the models →

Hardware watch

Memory decides what fits; memory speed decides how fast it answers. Prices are shown only with what they mean and when they were checked.

All hardware →

Graphics memoryThe memory on a graphics card. A model has to fit in it to run at full speed. tiers

Mutinai records launch prices (MSRPThe price the manufacturer set at launch. What a shop charges today can be very different.) and, where a retail price has been checked, the date it was checked. It keeps no price history yet, so it shows no price trends.

Unified memoryOne pool of memory shared by the processor and graphics, as on Apple silicon, so large models can fit.

These chips share one pool of memory with the processor, so what fits is chosen when the machine is bought, not by the chip.

What the community is talking about

All contributions →

Community Runs and reviews written by members will appear here, kept separate from the sourced facts everywhere else on this page and never merged into them.

Not collected yet

Contributions are not open yet

This preview has no sign-in, so nothing here could have been written by a member: the accounts, runs and reviews in its database are fictional sample content, and are left out rather than dressed up as activity. Real runs, reviews and discussion appear here once contributions open.

Why run models yourself

  • Your data stays put

    Prompts, code and documents stay on the machine you ran them on.

    How local models work →

  • No account, no per-token bill

    Download the weights once and run them on a laptop, one graphics card, a Mac or a server.

    What can I run? →

  • Nobody can take it back

    The weights are files you keep. Nobody can withdraw, reprice or quietly change them.

    What “open” means →