On August 13, Alibaba's Qwen team published the weights of Qwen3.8-2.4T-A95B — the 2.4-trillion-parameter mixture-of-experts model that powers the Qwen3.8-Max flagship. It is the first time a Max-level model has been opened, and the timing is no accident: the same night Grok 4.6 and DeepSeek V4 Pro landed, three frontier labs converging on one bet — long-horizon autonomous agents. But the open-source release may change the game more than the benchmarks do. Within hours, Unsloth had compressed the 4.9TB release to 397GB, meaning a single workstation with 410GB of combined RAM and VRAM can now run a frontier-class model locally.
What was actually released
- 2.4T total parameters, 95B active per token (MoE), built on the Qwen3.5 architecture
- Native 262,144-token context, extendable to ~1M tokens
- Gains concentrated in coding, office work, research, and long-horizon agent tasks
- Reasoning mode is mandatory — output starts with
<think>…</think>and cannot be disabled - Deployable via SGLang, vLLM, or TokenSpeed, each requiring the Qwen3.8-specific recipe
- A smaller Qwen3.8-27B is already announced
Weights are live on Hugging Face and ModelScope.
Benchmarks: strong where agents live, weaker on classic coding
The results are split. Wins: PaperBench 93.0 (ahead of GPT-5.6 Sol, Fable 5, and Claude Opus 4.8), OSWorld-Verified 86.1 (first among all evaluated models on computer operation), and 91.5 on the parametric CAD benchmark (above Fable 5, GPT-5.6 Sol, and Gemini 3.1 Pro). Losses: SWE-bench Pro 67.7 vs Fable 5's 80.0 and Opus 4.8's 69.2, TerminalBench 2.1 86.6 vs GPT-5.6 Sol's 88.8, and VideoMME v2 below GPT-5.6 Sol.
What Alibaba actually tested it on is more telling than any leaderboard: 16 days of continuous autonomous coding (building its own self-evolving harness, collecting community requirements, dispatching issues, generating code, verifying, and self-repairing), independently reproducing and improving a research paper, competing in real algorithm contests, and running a virtual e-commerce company across 2,000+ interaction rounds. Evaluation is shifting from one-shot answers to weeks-long autonomous work.
The structural change: open source just caught the frontier
For years, open weights meant "last year's model" — useful for fine-tuning, never for the frontier. Qwen3.8-2.4T breaks that pattern: the flagship itself is now reproducible, and the gap to the best closed models is measured in benchmark points, not generations. Compression is compounding the effect — Unsloth's dynamic 1-bit layered selective quantization cut the footprint 91%, and the "runs on my workstation" threshold keeps falling. The open ecosystem is no longer catching up; it's compressing the frontier into a deployable form.
What it means for the industry
- Self-hosting becomes real for regulated industries. Banks, governments, and healthcare can now run frontier-adjacent capability behind their own firewalls; data sovereignty and long-horizon agent workloads no longer force a compromise.
- Inference engines become the moat. With weights public, value shifts to the recipes — whoever optimizes SGLang/vLLM/TokenSpeed best controls distribution.
- Pricing pressure on closed APIs. At ¥12/$2 per million input tokens, marathon agent runs (thousands of turns) get dramatically cheaper — and that is exactly the workload everyone is racing toward.
- The leaderboard battle is redefining. SWE-bench alone no longer tells the story; research reproduction (PaperBench) and computer operation (OSWorld) are the new battlegrounds.
What to do next
- Building agent workflows? Start with Qwen3.8-Max on the API — the per-token pricing makes long-horizon experiments affordable.
- Running infrastructure? Test the Unsloth 397GB build; 410GB of RAM+VRAM is a beefy but purchasable single machine.
- Choosing a model? Re-run your own SWE-style tasks — this model wins on research reproduction and computer operation, not on every coding benchmark.
- Watch the 27B release; small-model follow-ups are where adoption compounds.
For context on where this fits: our 2026 AI trends overview tracks the agent-in-production shift, the DeepSeek V4 Pro launch analysis covers the same night's closed-side bet on tool-first execution, and the 2026 model selection guide helps you compare models by lane.
The companion Qwen 3.8-2.4T Open Weights: the Agent-First Recipe digs into the architecture recipe, infra and agent-first post-training of the same model.
