AI/TLDR

Qwen (Alibaba) · 2026-04-20 · major

Qwen3.6-Max-Preview — Alibaba's 1T-Parameter Coding Flagship, #5 on SWE-bench Pro

Alibaba's new API-only Qwen3.6-Max-Preview is a 1-trillion-parameter MoE model with 262k context, scoring #5 on SWE-bench Pro Public and #6 on Terminal-Bench 2.0 — free to try on Qwen Studio and Alibaba Cloud Bailian.

QwenLM Qwen3.6 GitHub repository — the family containing the new Max-Preview flagship model

Alibaba's new 1T-parameter MoE flagship scores #5 on SWE-bench Pro Public and leads six coding benchmarks over Qwen3.6-Plus.

Key specs

Parameters~1T (MoE)
Context window262,144 tokens
Swe bench pro public57.30 (rank 5/27)
Terminal bench 2.065.40 (rank 6/34)
Skills bench vs plus+9.9 pts
Sci code vs plus+10.8 pts

What is it?

Qwen3.6-Max-Preview is the new API-only flagship of the Qwen3.6 series, released April 20, 2026 by Alibaba's Qwen team. It is described as a preview of the next-generation Qwen model — larger and more capable than the Qwen3.6-Plus that recently shipped — with the biggest gains in agentic coding, world knowledge retention, and instruction-following. It is currently free on Qwen Studio and callable via Alibaba Cloud Bailian API under the model name 'qwen3.6-max-preview'.

How does it work?

The model uses a sparse Mixture-of-Experts architecture (~1 trillion total parameters, 262k-token context, 8,192 max output tokens). Benchmark improvements over Qwen3.6-Plus: SkillsBench +9.9 pts, SciCode +10.8 pts, NL2Repo +5.0 pts, Terminal-Bench 2.0 +3.8 pts. On the Artificial Analysis Intelligence Index it scores 52 and ranks #2 out of 201 models (median peer: 14). SWE-bench Pro Public: 57.30 (rank 5/27); Terminal-Bench 2.0: 65.40 (rank 6/34). World knowledge benchmarks also improved: SuperGPQA +2.3 pts, QwenChineseBench +5.3 pts.

Why does it matter?

The Qwen3.6 open-weight series already set a high bar for coding agents; the Max-Preview shows where the closed-API tier now sits — well above the Plus tier. It is currently free to access, making it a direct comparison point against Anthropic, OpenAI, and Google for anyone benchmarking coding pipelines or agentic tasks.

Who is it for?

Teams building or evaluating coding agents, agentic pipelines, and complex instruction-following systems.

Try it

Model name: qwen3.6-max-preview on Alibaba Cloud Bailian or Qwen Studio (chat.qwen.ai)

Sources · 4 outlets

Tags

  • qwen
  • alibaba
  • model
  • coding
  • agents
  • swe-bench
  • moe
  • api
  • agentic-coding
  • instruction-following

← All releases · Learn AI