Parameters
27B dense
Context
100K+
License
Apache 2.0
Vision
No

Qwen3.6 27B is the density king of the 24GB class: strong general quality, excellent code, first-tier multilingual output (including Italian, which many open models handle poorly), and a hybrid thinking mode you can toggle per request for harder reasoning. At Q4 it fits a single used RTX 3090 with room for real context, which makes it the default recommendation for a first serious local deployment.

VRAM by quantization level

QuantWeightsFits on
Q4_K_M ~16 GB 24GB comfortably; 16GB tight, short context
Q5_K_M ~19 GB 24GB with moderate context
Q8_0 ~29 GB 32GB (RTX 5090) or 48GB; 100K context fits on 32GB with Q8 KV-cache

Weights only — add KV-cache (grows with context and concurrency) and ~1–2GB runtime overhead. Formula and cache math in the VRAM guide.

Quick start

$ ollama run qwen3.6:27b # default Q4_K_M quant
$ ollama run qwen3.6:27b-q8_0 # near-lossless, needs 32GB+

LM Studio: search “Qwen3.6 27B”, pick the quant with the green fits-your-VRAM badge (MLX build on Macs). Production: vLLM with the AWQ build for multi-user serving. Raise the context — Ollama defaults num_ctx low; set 32768+ in a Modelfile or the thinking mode will truncate its own reasoning.

Expected performance

HardwareGeneration speed
RTX 3090 (Q4) ~28–35 tok/s
RTX 4090 (Q4) ~35–45 tok/s
RTX 5090 (Q8) ~35–45 tok/s
Mac M-series Max/Ultra ~12–20 tok/s

Indicative single-user figures (llama.cpp/Ollama class runtimes); multi-user serving via vLLM multiplies total throughput 10–20× through batching.

Tips & gotchas

  • Thinking mode multiplies output tokens — great for math/planning, wasteful for summaries. Toggle it per request, don't leave it always on.
  • For agents and JSON output, use Q5+ — structured output degrades before prose under quantization.
  • Enable Q8 KV-cache to double usable context on the same card.
  • Fully permissive — commercial use, modification and redistribution allowed.

FAQ

Can I run Qwen3.6 27B on 16GB of VRAM?

Yes at Q4 with short context (4–8K), but 24GB is the comfortable tier: full quality quant plus working context.

Is it good in Italian?

Among the best open models for Italian output quality — a key reason it's our default recommendation for Italian teams.

Thinking mode on or off?

Off by default; on for math, code architecture and multi-step planning. It costs tokens and latency.