Alibaba Releases Qwen3.6-Max-Preview — Closed-Weights Flagship Tops SWE-bench Pro at 57.3% (April 20, 2026)
Alibaba's Qwen team shipped Qwen3.6-Max-Preview on April 20, 2026, a hosted-only flagship that claims #1 spots on six coding and agentic benchmarks — including 57.3% on SWE-bench Pro — while quietly dropping the open-weight release Qwen has been famous for.
Alibaba's Qwen team released Qwen3.6-Max-Preview on , a hosted, proprietary flagship that claims the #1 score on six coding and agentic benchmarks — led by 57.3% on SWE-bench Pro — and undercuts Anthropic's Claude Opus 4.7 on price. For the first time in the Qwen series' history, the new flagship ships closed-weights only, available exclusively via Alibaba Cloud's Dashscope and Bailian platforms.
What Happened
Alibaba previewed Qwen3.6-Max on the official Qwen blog on April 20, calling the release "smarter, sharper, still evolving." The model is the largest in the Qwen3.6 family — sized above Qwen3.6-Plus, which set its own SWE-bench record in late March — and is the first Qwen "Max" tier whose weights are not slated for any open release. Per the announcement, Max-Preview is hosted on Alibaba Cloud Model Studio with API endpoints compatible with both the OpenAI and Anthropic specifications, accessible through the company's dashscope and bailian services.
The model offers a 260,000-token context window and a new preserve_thinking flag that keeps internal reasoning state across turns — a feature Qwen positions specifically at multi-turn agentic workflows where chain-of-thought continuity matters. Compared with its predecessor Qwen3.6-Plus, Alibaba reports gains of +9.9 points on SkillsBench, +10.8 on SciCode, and +3.8 on Terminal-Bench 2.0, plus +2.3 on SuperGPQA world-knowledge and +5.3 on QwenChineseBench.
Key Details
- Six benchmark #1s — Qwen claims top scores on SWE-bench Pro (57.3%), Terminal-Bench 2.0, SkillsBench, QwenClawBench, QwenWebBench, and SciCode at launch.
- Closed-weights pivot — unlike Qwen3.6-Plus, Qwen3.6-32B-A3B, and Qwen3.6-27B (all of which shipped open weights earlier this year), Max-Preview is hosted-only.
- API pricing — listed at $1.30 input / $7.80 output per million tokens on Dashscope, undercutting Anthropic Claude Opus 4.7 ($15 / $75) and OpenAI's GPT-5.5 Pro tier on output tokens by a wide margin.
- API compatibility — Alibaba's compatible-mode endpoint accepts both OpenAI Chat Completions and Anthropic Messages API payloads, easing migration for existing agent codebases.
- Agentic features — first-class function calling, the new
preserve_thinkingflag for multi-turn reasoning, and Anthropic-style tool-use formatting with the ToolcallFormatIFBench score up 2.8 points over Qwen3.6-Plus. - Branding — the GitHub organisation has rebranded the umbrella repo to QwenLM/Qwen3.6; Max-Preview itself does not include weight artifacts.
What Developers and Users Are Saying
The Hacker News thread on the release became the most-discussed Qwen post of 2026 within hours. The dominant theme: developers welcome the SWE-bench number but are frustrated by the closed-weights direction. One commenter (seplite) wrote that “like with the release of Qwen3.6-Plus, this model also isn't released for local use,” quoting Qwen's own line that “Qwen3.6-Max-Preview is the hosted proprietary model available via Alibaba Cloud Model Studio.” Others pushed back that the Max tier was always API-only, and noted that the smaller MoE Qwen3.6-32B-A3B and dense Qwen3.6-27B models — both released earlier this year — remain freely downloadable.
On price, the comparison with Anthropic dominated. One top-voted reply argued that Codex 5.4 / Claude Opus 4.7 are still the SOTA for users where “money is no object,” but that for “many to most people, value vs. relative quality are huge levers” — a clear opening for Qwen3.6-Max at roughly one-tenth the per-token output cost of Opus 4.7. Independent reviewers at TokenMix and BuildFastWithAI have so far reproduced the SWE-bench numbers within 1–2 points, while flagging the usual caveat that vendor-published benchmarks tend to omit the frameworks where a model fares poorly.
What This Means for Developers
For agent-builders, Qwen3.6-Max-Preview is suddenly the cheapest model with a credible claim to top-of-table SWE-bench Pro performance — and with OpenAI- and Anthropic-compatible endpoints, swapping it into an existing Claude or GPT pipeline takes a base-URL change and an API key. Teams already using Qwen's open-weight 27B/32B for self-hosted inference can keep those for cost-sensitive paths and route premium agent runs to Max via the same SDK.
The closed-weights pivot also matters for the open-weight community. Qwen has been the most reliable counterweight to OpenAI and Anthropic on weight releases since 2024; Max-Preview being hosted-only continues a trend that started with Qwen3.6-Plus and signals that Alibaba's frontier work will increasingly live behind an API. Self-hosters and on-prem buyers will need to keep an eye on whether the next Max tier returns to open weights, as Decrypt notes the company has not commented publicly on a long-term policy.
What's Next
Alibaba labelled the release a "preview," with a general-availability version expected within weeks. The official QwenLM/Qwen3.6 repository tracks SDK examples, benchmark configs, and an Anthropic-compatible client; the Dashscope and Bailian model cards list region-by-region rollout and rate-limit details. Independent third-party benchmarks from Artificial Analysis and SWE-bench are expected to publish reproductions in the coming days, and OpenRouter has indicated it will expose Qwen3.6-Max-Preview through its provider routing once Alibaba enables the necessary endpoints.
Sources
- Qwen Team Blog: Qwen3.6-Max-Preview — Smarter, Sharper, Still Evolving — primary source from Alibaba's Qwen team.
- QwenLM/Qwen3.6 on GitHub — official repository with SDK examples and changelog.
- Hacker News discussion thread — developer reactions, including commentary on the closed-weights pivot.
- Decrypt — Alibaba Drops Qwen 3.6 Max Preview — independent reporting on positioning vs. Claude and GPT.
- TokenMix — Qwen3.6-Max-Preview Benchmark Review — independent reproduction of headline benchmarks.
- BuildFastWithAI — Qwen3.6-Max-Preview Review — additional independent benchmark notes and API walkthrough.
Stay up to date with Doolpa
Subscribe to Newsletter →