Qwen3.6-Plus: Towards Real World Agents

TL;DR

Qwen3.6-Plus is Alibaba’s closed, API-only flagship in the Qwen3.6 generation, hosted on Alibaba Cloud Model Studio with a 1M-token default context window and an “always-on” reasoning mode aimed at multi-turn agentic workflows. The launch foregrounds agentic coding (SWE-bench Verified 78.8, Terminal-Bench 2.0 ~61.6) and tool-using agents, with a new preserve_thinking API parameter that retains chain-of-thought across conversation turns to reduce redundant reasoning in long agent loops. It exposes both OpenAI- and Anthropic-protocol-compatible endpoints, advertising drop-in use from Claude Code, OpenClaw, Qwen Code, Kilo Code, Cline, and OpenCode. Smaller open-weight variants in the 3.6 family (e.g. 35B-A3B) are flagged for follow-up release. Relevant to Luma as the closed-source counterpart to the open Qwen3.5/3.6 line — same vendor, same generation, different deployment surface.

Key claims

Qwen3.6-Plus is API-only via Alibaba Cloud Model Studio (no public weights at launch); the post positions it as the production-tier hosted model in the Qwen3.6 family [§Intro].
Default 1M-token context window, exposed as the API baseline rather than an experimental toggle [§Intro].
The release introduces a preserve_thinking API parameter: when true, the model retains thinking content from all preceding turns; default is false (only current-turn thinking is kept) [§API parameter description].
The stated motivation for preserve_thinking is maintaining decision consistency across long agentic trajectories and reducing redundant re-derivation of reasoning, which can lower aggregate token consumption [§API parameter description].
API supports both OpenAI- and Anthropic-protocol-compatible endpoints; the blog enumerates Claude Code, OpenClaw, Qwen Code, Kilo Code, Cline, and OpenCode as supported third-party coding harnesses [§Coding integrations].
Qwen3.6-Plus claims SWE-bench Verified ≈78.8, within ~2 points of Claude Opus 4.5/4.6 on the same benchmark, and Terminal-Bench 2.0 ≈61.6 where it leads Claude Opus on the open report [§Agentic Coding].
For general-purpose agents and tool use, the post claims top results on multiple long-horizon planning tasks and leadership across tool-calling benchmarks (no per-benchmark numbers quoted in the visible body) [§General Agents & Tool Use].
Multimodal claims center on document understanding, physical-world visual analysis, video reasoning, and visual coding, with the post framing these as workflow-oriented multimodal capabilities rather than isolated task-benchmark gains [§Multimodal].
Qwen Code OAuth provides 1,000 free Qwen3.6-Plus API calls per day per signed-in user [§Free Access].
The post sequences Qwen3.6-Plus after the Qwen3.5 series (Feb 2026) and frames it as “a massive capability upgrade over its predecessor”, with smaller open-source variants in the Qwen3.6 family expected to follow within days [§Intro].

Method

The blog is a launch announcement, not a technical report. Architecture details are not disclosed beyond the framing of “sparse mixture-of-experts with always-on reasoning” picked up by third-party coverage; the post itself only documents the API surface, the preserve_thinking parameter, the protocol-compatibility surface (OpenAI + Anthropic), and the supported coding-harness integrations. The headline new mechanism is preserve_thinking: a per-request boolean that, when true, prepends prior-turn thinking spans into the input context for subsequent assistant turns, rather than discarding them as is standard for hosted reasoning models. The post is explicit that the parameter is recommended specifically for agentic tasks, not for chat — its win is conditional on the agent needing to refer back to its own past reasoning steps.

Results

Numbers are vendor-self-reported on the launch post; not independently reproduced.

Coding / SWE. SWE-bench Verified ≈78.8 (within ~2 points of Claude Opus 4.5/4.6’s ~80.8–80.9). Terminal-Bench 2.0 ≈61.6, claimed to beat Claude Opus on this benchmark.
Context. 1M tokens default on the API; “256K used for some standardized benchmarks” per third-party coverage of the launch material.
Multimodal. Qualitative claims of substantial breakthroughs in complex document understanding, physical-world visual analysis, video reasoning, and visual coding; no per-benchmark numbers in the visible blog body.
General agents. Qualitative claims of top results on “multiple challenging long-horizon planning tasks” and leadership across tool-calling benchmarks; no per-benchmark numbers in the visible blog body.
Multilinguality and STEM. Qualitative claim of “new records” in difficult STEM reasoning, ultra-long-context information extraction, and broad multilingual adaptation; no per-benchmark numbers in the visible blog body.
Free-access surface. Qwen Code OAuth grants 1,000 free Qwen3.6-Plus calls per day per user.

Why it’s interesting

This is the closed-frontier successor in the Qwen line that Qwen3.5: Towards Native Multimodal Agents (open, Apache 2.0, 397B-A17B) sits beneath — same vendor, same generation, different deployment surface. It’s the open–closed boundary of the Qwen3.x family made explicit: Qwen3.6-Plus is API-only with smaller open variants promised to follow. On the agentic-coding axis it slots in alongside Introducing GPT-5.3-Codex and Step 3.5 Flash as a 2026-Q1/Q2 closed/hosted datapoint reporting on the SWE-bench Verified + Terminal-Bench 2.0 quadruple that Agentic Software Engineering tracks — Qwen’s SWE-Verified ~78.8 trails Claude Opus 4.6 by ~2 points but reportedly leads on Terminal-Bench 2.0 in the launch comparison. The preserve_thinking parameter is the new design lever to watch: it turns “what to do with prior-turn thinking spans in a hosted reasoning model” into a documented, per-request API control, which is one specific answer to the chain-of-thought-persistence question raised by long-horizon agentic systems in MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling and Reasoning RL. Direct sibling to closed launch announcements Nano Banana 2 / Gemini 3.1 Flash Image (Google product announcement) (Gemini 3.1 Flash Image) and GPT-5.4 Thinking and GPT-5.4 Pro rolling out in ChatGPT (product announcement) (GPT-5.4) — same March–April 2026 release window of vendor-blog launches with strong benchmark claims and no architecture disclosure.