AI ranking

Local AI — code

Open-weights models (qwen, deepseek, llama, mistral…) of 35B parameters or fewer, assessed on code (SWE-bench, Aider). Models that run locally on a consumer GPU or an Apple Silicon Mac.

Ranking updated on 02 November 2026 at 11:47

Visual top 10

  1. 01 Qwen3.8 27B 65.17
  2. 02 Qwen3.6 27B 52.29
  3. 03 Qwen3.5 27B 36.96
  4. 04 QwQ 32B 34.25
  5. 05 Qwen2.5 Coder 32B Instruct 33.18
  6. 06 Gemma 3 27B It 32.32
  7. 07 Qwen3.5 35B A3b 24.15
Local limit: ≈35B max

Models beyond roughly 35 billion parameters are not ranked in the local categories. Le Recul favours models that can genuinely be run locally, without heavy professional infrastructure.

Le Recul comparator

Who wins on what

Tick at least two models in the ranking table, then come back here.

# Model Vendor Score Assessment
01 Alibaba 65.17 / 100 Partial
02 Alibaba 52.29 / 100 Partial
03 Downward trend Alibaba 36.96 / 100 Partial
04 Alibaba 34.25 / 100 Partial
05 Alibaba 33.18 / 100 Partial
06 Google 32.32 / 100 Reliable
07 Alibaba 24.15 / 100 Partial

Indicative score, weighted from public sources. Transparent methodology. See the AI ranking page for the full methodology and the other categories.

Local AI ranking for coding

This view gathers open-weight models compact enough to run locally — typically up to around 35 billion parameters — assessed on code: fixing real bugs (SWE-bench), editing files (Aider), long reasoning, reliability, cost and speed.

The appeal is concrete: coding with a local model means keeping your code on your own machine, working offline and not paying per token. For code under NDA or subject to confidentiality constraints, it is sometimes the only acceptable option.

The gap has to be kept in mind, though. Compact local models remain behind the best cloud models on long tasks and complex code bases. They are often very good on isolated functions, tests or review, and more fragile as soon as a wide context has to be held.

What this ranking is for

This view helps you choose a model that runs on a consumer GPU or an Apple Silicon Mac to assist development without sending your code to a third party.

It also helps calibrate expectations: aim at the tasks where these models are genuinely effective rather than handing them a full rewrite and concluding too quickly that they are useless.

How to read the ranking

  • SWE-bench: ability to fix real bugs in existing code.
  • Aider: reliable file editing, output format respected.
  • Long reasoning, decisive on multi-step tasks.
  • Reliability and consistency between two identical runs.
  • Hardware cost and local execution speed.

FAQ — Local AI — code

What hardware do you need to code with a local AI?

The models in this view target a consumer GPU or a recent Apple Silicon Mac. Available memory is the main limiting factor, more than raw power.

Can a local AI replace a cloud assistant for coding?

On targeted tasks — an isolated function, a test, a review — often yes. On long tasks involving a wide code base, the gap with the best cloud models remains clear.

Why is the list short?

Because the filter is twofold: an open model that can be run locally, and assessed on code by public sources. Few models meet both conditions with enough coverage.

What is the difference with the Code ranking?

The Code ranking covers every model, cloud included. This view keeps only those you can run yourself.

How often is it updated?

About every 25 hours, automatically. The update date is shown at the top of the page.

See the full FAQ →

Related rankings