Qwen 3.6-27B Released — Dense 27B Leads Agentic Coding, 40 tok/s on RTX 3090 [April 2026]

Alibaba's Qwen Team released Qwen 3.6-27B on April 22, 2026 — the first Dense (non-MoE) open-weight model in the Qwen 3.6 family. It's published on Hugging Face and ModelScope under Apache 2.0, allowing commercial use and fine-tuning, with substantial improvements aimed at agentic coding.

Headline numbers from the official announcement:

The model reportedly beats both its predecessor Qwen 3.5-27B and the much larger Qwen 3.5-397B-A17B MoE on multiple tasks. Matching Claude 4.5 Opus on Terminal-Bench 2.0 with a 27B open model is the headline.

Despite being 27B Dense, Qwen 3.6-27B introduces a "Thinking Preservation" mechanism (per the official blog) and combines Gated DeltaNet linear attention with traditional self-attention in a hybrid architecture. Native context is 262,144 tokens, extensible to roughly 1,010,000. The model also accepts text, image, and video inputs.

Our existing posts on Qwen 3.5-9B (Complete Guide / Claude replacement migration guide) cover the lighter sibling, which is still a strong choice. Here's how to decide between the two:

Picking guidance: - Cost-first, run on a Mac mini → Qwen 3.5-9B - Agentic coding / dev productivity is the core need → Qwen 3.6-27B - Already have a 24GB+ GPU server / serious DX investment → Qwen 3.6-27B - Need it on a laptop / 8GB Mac → Qwen 3.5-9B - Need multimodal → Qwen 3.6-27B (3.5-9B is text-focused) A hybrid setup — Qwen 3.5-9B for routine queries and Qwen 3.6-27B only when coding judgment is required — is often the most cost-balanced choice.

Reported runs show ~40 tokens/sec on an RTX 3090 24GB with Q4_K_M quantization, passing 10 of 10 functional tests. Hitting practical throughput on a workstation-class consumer GPU makes this 27B Dense open model genuinely deployable for SMBs. Apple Silicon (M3 Max 64GB / M4 Max-class) is also viable for quantized inference.

Qwen 3.6-27B is now a candidate local backend for OpenClaw (see OpenClaw 2026.4.23 release notes). It pairs well with internal-DX projects where confidential data shouldn't leave a private network but agentic coding is still required, and Apache 2.0 fine-tuning is a practical bonus. For deployment guidance, see AI Consulting.

Q1: Worth upgrading from Qwen 3.5-9B? A: For coding-heavy workloads, likely yes. Note that 9B → 27B raises VRAM requirements significantly. Q2: Does it run on Mac mini? A: M4 Pro 32GB+ can run quantized Qwen 3.6-27B usefully. For M2 8GB-class, prefer Qwen 3.5-9B or Gemma 4 E4B. Q3: Is commercial use allowed? A: Apache 2.0 — commercial use, modification, and redistribution are all permitted.

- Qwen 3.6-27B Official Blog (Qwen team) - Hugging Face: Qwen/Qwen3.6-27B - Qwen Team Releases Qwen 3.6-27B (MarkTechPost, 2026/4/22) - Qwen 3.6-27B on RTX 3090 (Phemex News)