GitHub radar
Qwen3.6: Alibaba's multimodal coding model on Ollama
Alibaba's Qwen3.6-27B scores 77.2% on SWE-bench Verified and runs locally via Ollama — 262K context, multimodal input, and thinking preservation across agent steps.
Qwen3.6 is a multimodal model from Alibaba's Qwen team designed for agentic coding — it processes text, images, and video in a single session. The 27B version scores 77.2% on SWE-bench Verified, placing it among the most capable models that can actually run on consumer hardware. Context window is 262,144 tokens natively, meaning a full repository fits in one session. The new 'thinking preservation' feature lets the model retain its reasoning chain from previous messages, so an agent doesn't restart its logic on each tool call. Available on Ollama in two sizes: 27B at 18 GB and a 35B MoE variant with only 3B active parameters at inference time (23 GB).
Why a vibe-coder should care
For vibe coders who need a local model for real coding work — not just auto-complete — Qwen3.6-27B is currently one of the strongest options that fits on consumer hardware. The thinking preservation feature directly improves multi-step agent workflows by keeping reasoning context between tool calls. Honest take: it handles straightforward to moderately complex coding tasks well; for the most demanding multi-codebase agent work, frontier cloud APIs still have the edge.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install Qwen3.6 from Alibaba locally: ollama run qwen3.6 — pick the 27b size, then show me how to give it a coding task with a screenshot attached
A Mac with 32 GB of RAM or a GPU with 24 GB of VRAM — the 27B version takes about 18 GB
Open on Hugging Face▌ More finds