Which Qwen 3.6 variant should I run locally for coding?

Dense Qwen 3.6 27B is the recommendation from hands-on testing by Quesma’s Piotr Migdał (June 29, 2026). The MoE Qwen 3.6 35B A3B is roughly 3× faster on Apple Silicon but produced lower-quality agent output — e.g. ignoring package.json instructions in an OpenCode minesweeper task. Prefer 27B when quality matters; use 35B A3B when speed dominates and tasks are simple.

How do I run Qwen 3.6 27B with llama.cpp?

Pull unsloth/Qwen3.6-27B-MTP-GGUF:Q8_0 from Hugging Face, then run llama-server with -hf, --spec-type draft-mtp, -ngl 999, -fa on, -c 65536, and --port 8080. That exposes an OpenAI-compatible API at http://127.0.0.1:8080/v1 for OpenCode, Pi, or Hermes. Native context is 256k; 64k is a practical default for RAM.

How fast is Qwen 3.6 27B on Apple Silicon?

On a MacBook Pro Max M5 with 128GB RAM, Migdał measured ~18 tok/s with llama.cpp Q8, ~32 tok/s with multi-token prediction (MTP), using ~41–42GB RAM. The 35B A3B MoE hit ~93–105 tok/s at ~44–45GB. An HN user reported ~50 tok/s on RTX 5090 at Q6_K with 123k context (~28/32GB VRAM) via LM Studio.

How does Qwen 3.6 27B compare to Gemma 4 31B locally?

Artificial Analysis Intelligence Index scores cited in the Quesma post place Gemma 4 31B at 29 (~late 2024 tier), Qwen 3.6 35B A3B at 32 (~early 2025), Qwen 3.6 27B at 37 (~mid 2025), and DeepSeek V4 Flash quant at 40 (~late 2025). Community sentiment and coding smoke tests in the post favor 27B over Gemma 4 31B for local agent work — but run your own evals on your hardware.

Can Qwen 3.6 replace cloud models after the Fable 5 ban?

For bounded local tasks — landing pages, small apps, prose, offline sensitive data — yes, as a practical daily driver. It does not match frontier closed models on hard agentic benchmarks. Post-Fable teams tier workloads: local Qwen/GLM for private and volume work, cloud or API for peak reasoning. See Fable alternatives and GLM-5.2 local guides for the next step up.

Should I use Ollama or llama.cpp for Qwen 3.6?

Migdał recommends llama.cpp directly for control, MTP support, and layer offload tuning — and argues against Ollama on ethical grounds in the original post. explainx.ai’s read: llama.cpp or MLX on Mac for power users; Ollama remains the lowest-friction entry in our local AI stack guide if you prioritize one-command setup over maximum throughput.

← Back to blog

explainx / blog

Qwen 3.6 27B Local Dev Guide: llama.cpp, OpenCode, and Why Dense Beats MoE

Quesma’s Piotr Migdał benchmarks Qwen 3.6 27B on Apple Silicon — 32 tok/s with MTP, OpenCode minesweeper wins, AA score 37. llama.cpp setup, vs 35B MoE and Gemma 4 31B.

Jun 29, 2026·8 min read·Yash Thakker

QwenLocal LLMllama.cppOpenCodeAgentic CodingOpen Source AI

Qwen 3.6 27B Local Dev Guide: llama.cpp, OpenCode, and Why Dense Beats MoE

Related posts

VibeThinker 3B: A 3-Billion Parameter Model That Matches Opus 4.5 Performance

Gemma 4 12B: Multimodal Local AI Guide 2026

Agents-A1: InternScience 35B MoE Agent Model — Long-Horizon Search, GAIA 96, and vLLM Setup