Best Local LLM for Coding (2026): Ranked by RAM

Which Local LLM Is Best for Coding? The Tier List

Your RAM (or VRAM) is the hard ceiling, so the ranking below is organized by what actually fits. Benchmark figures shown are the ones published in each vendor's own model card and confirmed in the raw source; where no confirmed figure exists, we say so instead of inventing one. Every ollama run command was probed against the Ollama registry (HTTP 200) on 2026-06-12.

ollama run qwen3.6:27b

ollama run devstral-small-2:24b

ollama run qwen3-coder:30b

ollama run gpt-oss:20b

ollama run qwen2.5-coder:7b

ollama run qwen2.5-coder:1.5b

ollama run gpt-oss:120b

ollama run devstral-2:123b

Sources: Qwen3.6-27B model card (77.2 SWE-bench Verified, 59.3 Terminal-Bench 2.0), Devstral Small 2 model card (68.0% SWE-bench Verified; Devstral 2 123B at 72.2%), aider polyglot leaderboard (gpt-oss-120b 41.8%). SWE-bench figures are vendor self-reported on their model cards; aider scores are independently run.

What Changed in 2026: the Local Coding Gap Collapsed

A year ago the standard answer was Qwen2.5-Coder 32B — which scores just 16.4% on the aider polyglot benchmark. The 2026 generation rewrote that completely: Devstral Small 2 posts a 68.0% SWE-bench Verified score in a 15GB download, and Qwen3.6 27B reports 77.2 — numbers that were frontier-cloud territory in 2025. If your "best local coding model" knowledge dates from the Qwen2.5-Coder era, every recommendation you remember is out of date.

Two caveats keep cloud models ahead for some work: very long agentic sessions (local context windows fill up faster in practice) and the biggest open-weight coders (DeepSeek, Kimi K2) are 600B-1T-parameter models that no consumer machine runs — open-weight is not the same as locally runnable. For a deeper comparison, see our local vs cloud flagship benchmark breakdown.

How Do I Use a Local LLM in VS Code? (continue.dev)

continue.dev is the most popular open-source Copilot alternative for VS Code and JetBrains. Configuration now lives in ~/.continue/config.yaml (the old config.json is deprecated). A complete local setup — chat on a strong model, autocomplete on a fast one:

models:
  - name: Qwen3.6 27B
    provider: ollama
    model: qwen3.6:27b

  - name: Autocomplete
    provider: ollama
    model: qwen2.5-coder:1.5b
    roles:
      - autocomplete

Qwen2.5-Coder 1.5B is continue.dev's officially recommended autocomplete model. Their docs also warn against using thinking/reasoning models for autocomplete — completions need to arrive in milliseconds, not after a reasoning chain.

How Do I Use aider with Ollama?

aider is the terminal-native AI pair programmer. Point it at your local Ollama server:

export OLLAMA_API_BASE=http://127.0.0.1:11434
aider --model ollama_chat/qwen3.6:27b

Use the ollama_chat/ prefix — aider's docs state it is recommended over plain ollama/. The critical gotcha: Ollama defaults to a 2K context window and silently discards anything beyond it (aider's own warning). Coding needs far more. Raise it server-side:

OLLAMA_CONTEXT_LENGTH=32768 ollama serve

How Do I Use a Local LLM in Zed?

Zed auto-discovers every model Ollama has pulled — pull a model, then pick it from the Agent panel's model dropdown. One thing worth overriding: Zed defaults local models to a 4096-token context. In settings.json:

{
  "language_models": {
    "ollama": {
      "api_url": "http://localhost:11434",
      "available_models": [
        {
          "name": "qwen3.6:27b",
          "display_name": "Qwen3.6 27B",
          "max_tokens": 32768,
          "supports_tools": true
        }
      ]
    }
  }
}

Frequently Asked Questions