AI ranking

Local AI — agentic

Open-weights and local models (Qwen, DeepSeek, Gemma, Llama, Phi, Mistral… 35B or fewer) suited to agentic use: models present in both chat and code. These are models usable inside agentic systems, not applications or scaffolds.

Ranking updated on 02 November 2026 at 11:47

Visual top 10

  1. 01 Qwen3.8 27B 85.82
  2. 02 Qwen3.5 27B 82.56
  3. 03 Qwen3.5 35B A3b 77.24
  4. 04 Qwen3.6 27B 66.58
  5. 05 QwQ 32B 52.48
  6. 06 Gemma 3 27B It 48.89
  7. 07 Qwen2.5 Coder 32B Instruct 31.54
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 85.82 / 100 Partial
02 Alibaba 82.56 / 100 Reliable
03 Alibaba 77.24 / 100 Reliable
04 Alibaba 66.58 / 100 Reliable
05 Alibaba 52.48 / 100 Reliable
06 Google 48.89 / 100 Reliable
07 Alibaba 31.54 / 100 Reliable

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 agentic use

This view gathers locally runnable open-weight models that hold up in both conversation and code — the dual competence expected of a model meant to drive an agent. These are models usable inside an agentic system, not agentic applications.

Running an agent locally answers a precise need: keeping data on your own machine, avoiding per-token billing on long, repeated tasks, and not depending on a third-party service staying available. On agentic loops that consume many calls, the economic argument quickly becomes decisive.

The requirement is higher than for chat, though. An agent chains dozens of steps: a small weakness propagates and amplifies at every turn. Consistency therefore matters more here than peak performance, and that is often where compact models show their limits.

What this ranking is for

This view helps you choose a local model able to hold an agentic loop without drifting after a few steps.

It also calls for stricter supervision: a badly framed local agent fails silently, without the safety net of a provider that cuts it off or flags the anomaly.

How to read the ranking

  • Simultaneous presence in chat and in code (entry condition for the view).
  • Overall quality and reasoning, the basis for holding up on a long task.
  • Consistency between runs, decisive in an agentic loop.
  • Local execution cost and speed, across a high volume of calls.
  • Accessibility: weights available, formats and runtimes supported.

FAQ — Local AI — agentic

Can an AI agent really run locally?

Yes, on framed tasks and with a suitable model. The limits appear on long, loosely specified tasks, where drift sets in step after step.

What is the difference with the AI agents ranking?

The AI agents ranking assesses complete systems. This view only ranks the local models that could power them.

Why does consistency matter more than the raw score?

Because an agent chains many steps. A one-off error with no consequence in chat can derail a whole chain of actions.

Do you have to supervise a local agent?

Yes, at least as much as a cloud agent. Running locally removes the billing, not the risk of unwanted actions.

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