Qwen3.6-27B: A Dense 27B That Beats Its Own 397B MoE

Quick verdict

If you have one good GPU (RTX 5090, RTX 4090 with tight budget, or M3 Max/Ultra) and you want one local model that handles coding + research + general work, Qwen3.6-27B is the right default — and three months after release (checked July 2026), that still holds. The June-July wave of giant open models (GLM-5.2, Hy3, MiniMax M3) all target multi-hundred-GB rigs. The one release that does compete in this class is its own successor, Qwen3.8-27B (August 2026); see the comparison box above before you pick.

For multi-GPU rigs or H100-class hardware, Qwen3-Coder-Next (80B/3B active MoE) edges ahead on coding benchmarks. For frontier-class general capability with 1M context, DeepSeek V4-Flash is the upgrade. On a 16GB card, see the best coding LLMs for 16GB VRAM — Qwen3.6-27B's Q4 build is right at that boundary.

Specs at a glance

Why a 27B dense model beats a 397B MoE

The whole 2024-2025 narrative was that MoE wins by spending compute only where needed. Qwen3.6-27B challenges that for agentic coding workloads. Three reasons it works:

• 1.All-active compute every token. A 27B dense model applies all 27B parameters to every input token. Qwen3.5-Plus 397B MoE activates only 17B per token. For tasks where every step needs deep reasoning (multi-file refactors, debugging loops), all-active beats sparse-active.
• 2.Stronger inter-layer information flow. Dense transformers route information through every layer in a single integrated path. MoE creates multiple parallel paths that have to be reconciled — fine for facts and patterns, less ideal for stepwise reasoning.
• 3.Better training data. Qwen3.6-27B trained on a curated coding corpus with longer agentic-loop traces, while Qwen3.5-Plus prioritized broad knowledge. For coding benchmarks specifically, that data quality dominates parameter count.

Caveat: dense doesn't beat MoE everywhere. Qwen3.5-Plus 397B still wins on broad knowledge benchmarks (MMLU-Pro 88.4% vs Qwen3.6-27B's 81.7%), where parameter count and breadth matter more than reasoning depth. The right takeaway: pick architecture by workload, not by parameter count.

Hardware & setup

Upper bounds, not measurements: each generated token reads every weight once, so memory bandwidth (vendor spec sheets) divided by file size (from Ollama's qwen3.6 tag list, checked November 2, 2026) caps the speed. Real throughput lands below the ceiling, and further below it as the context fills.

Ollama (5-min install)

ollama pull qwen3.6:27b
ollama run qwen3.6:27b
# Or for Cursor/Aider integration:
ollama serve  # listens on :11434/v1

vLLM (production)

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen3.6-27B --max-model-len 131072 --port 8000

Benchmarks

When to pick Qwen3.6-27B

• ✓You have one good GPU and want one local model for coding + general work.
• ✓You want simpler deployment than MoE (predictable VRAM, no expert routing complexity).
• ✓You're on Apple Silicon — dense models work cleanly on M3 Max/Ultra Metal.
• ✓Your work fits in 128K context (most coding does).

Frequently asked questions