Qwen3.6-27B announcement: 27B dense beats Qwen3.5-397B-A17B on coding benchmarks

TL;DR

Qwen team announcement of Qwen3.6-27B — a 27B dense, natively multimodal LM/VLM positioned as the Qwen3.6-generation’s small-dense counterpart to the larger MoE Qwen3.6-35B-A3B and the closed Qwen3.6-Plus. The headline claim is that the 27B dense model outperforms the ~15× larger Qwen3.5-397B-A17B (397B total / 17B active) MoE on every reported coding benchmark — SWE-bench Verified 77.2 vs 76.2, SWE-bench Pro 53.5 vs 50.9, Terminal-Bench 2.0 59.3 vs 52.5, SkillsBench 48.2 vs 30.0 — and also surpasses peer-scale dense models by a wide margin. The card is a tweet, not a paper or HF release page, so no architecture, training recipe, or independent reproduction is available; numbers are vendor-self-reported.

Key claims

Qwen3.6-27B is a 27B-parameter dense model, native multimodal (vision + text), with both thinking and non-thinking modes in a single unified checkpoint — the same dual-mode property as the MoE Qwen3.6-35B-A3B sibling [tweet body].
Coding-benchmark numbers (vendor-reported): SWE-bench Verified 77.2, SWE-bench Pro 53.5, Terminal-Bench 2.0 59.3, SkillsBench 48.2 — all ahead of Qwen3.5-397B-A17B’s 76.2 / 50.9 / 52.5 / 30.0 at ~15× larger total parameters and ~6× larger active parameters [tweet body].
VLM capabilities span image and video alongside text: multimodal reasoning, document understanding, and visual question answering, in the same checkpoint as the LM [tweet body].
The model is positioned as surpassing “all peer-scale dense models by a wide margin” (no peer dense models named in the tweet) [tweet body].

Method

No method, training recipe, architecture, or release artifact is described in the announcement. The tweet is purely a benchmark-comparison post.

What can be inferred from the surrounding Qwen3.6 generation already filed: the MoE sibling Qwen3.6-35B-A3B Model Card uses a 3:1 Gated-DeltaNet:Gated-Attention hybrid-linear backbone with 256-expert MoE and a preserve_thinking API option; whether Qwen3.6-27B inherits the hybrid-linear backbone (now applied dense) or reverts to standard full attention is not stated. The closed Qwen3.6-Plus blog (Qwen3.6-Plus: Towards Real World Agents) had promised “smaller open variants would follow”; this looks like the second such variant after 35B-A3B, and the first dense one in the generation.

Results

All numbers are vendor-reported in the announcement tweet. No independent reproduction.

SWE-bench Verified. 77.2 vs Qwen3.5-397B-A17B 76.2 (+1.0). Also above the 35B-A3B sibling’s 73.4 Qwen3.6-35B-A3B Model Card.
SWE-bench Pro. 53.5 vs Qwen3.5-397B-A17B 50.9 (+2.6). Above the 35B-A3B sibling’s 49.5 and above all open-MoE peers on this page’s tracked Pareto (DeepSeek-V3.2 40.9, GLM-4.7 40.6 per Agentic Software Engineering §Key claims). Closing on closed-source GPT-5.3-Codex’s 56.8.
Terminal-Bench 2.0. 59.3 vs Qwen3.5-397B-A17B 52.5 (+6.8). Above 35B-A3B’s 51.5 and above Step 3.5 Flash’s 51.0.
SkillsBench. 48.2 vs Qwen3.5-397B-A17B 30.0 (+18.2). Largest gap in the reported set, suggesting a benchmark where the 27B is much better tuned than the older 397B-A17B flagship — but no description of what SkillsBench measures appears in the tweet.
VLM benchmarks. Mentioned categorically (multimodal reasoning, document understanding, VQA) but no numbers in the announcement.

Why it’s interesting

The headline — 27B dense beats 397B-total / 17B-active MoE on every coding benchmark in the table — is the strongest published datapoint yet that the active-parameter-efficiency advantage of MoE is not monotonic at the agentic-SWE Pareto, and that better training recipe at one generation can outweigh ~6× more active parameters at the prior generation. The Agentic Software Engineering cluster has been trending toward low-active-ratio MoE as the dominant architecture (Qwen3-Coder-Next 80B/3B, DeepSeek-V3.2 37B-active, GLM-4.7 32B-active, Step 3.5 Flash 11B-active) — Qwen3.6-27B is a dense outlier on that Pareto, in the same way IQuest-Coder-V1 Model Family is, but at a frontier-tier quality level. The Qwen3.6 generation now spans three open form factors: 27B dense (this tweet), 35B-A3B MoE (Qwen3.6-35B-A3B Model Card), and Plus-tier closed (Qwen3.6-Plus: Towards Real World Agents) — useful as a controlled triangle for whoever wants to study dense-vs-MoE under matched training data and recipe. One caveat the wiki should flag: this is a tweet with no technical report, no HF link, no architecture description, and an undefined benchmark (“SkillsBench”) where the gap is +18 points — the largest delta in the set sits on the least documented axis.