Verdict: Qwen 3.8 Max is Alibaba's most capable model to date and the first Qwen-Max-class model promised to ship open weights — a genuine inflection in the 2026 open-weight race. Its standalone strengths are long-horizon autonomous coding and general agent work, where its own published benchmarks put it ahead of GPT-5.6 Sol on five of seven agent tasks; but on the hardest frontier-coding bars (SWE-bench Pro, FrontierSWE, Humanity's Last Exam) Claude Fable 5 still wins by wide margins. Treat it as a serious frontier-tier option that undercuts US labs on price, not as a Fable-killer (Vendor claim, benchmarks self-reported; independent verification still pending on most scores).
What Is Qwen 3.8 Max and why it matters now
Qwen 3.8 Max is Alibaba's flagship AI model, built on a sparse Mixture-of-Experts (MoE) architecture with 2.4 trillion total parameters and 95 billion active per token. It succeeded Qwen 3.7-Max (released May 2026) and was officially released on August 3, 2026, after a paid preview that ran from July 19 under the model ID qwen3.8-max-preview. For the first time on a Max-tier Qwen, Alibaba has promised open weights, which it says will follow "next week" as of the August 3 release.
Why it matters: open weights at this scale would be the largest Apache-2.0-class open model ever shipped from a major lab. Until now, Alibaba kept every Max-tier model closed, serving it only through Alibaba Cloud Model Studio and Qwen Studio. It arrives on the heels of Moonshot AI's Kimi K3 (2.8T parameters, open weights promised for July 27) and Z.AI's GLM 5.2 — three Chinese trillion-parameter-class releases in barely a month, each claiming a place near the top of the global leaderboard. For builders outside China, Qwen 3.8 Max is the cheapest frontier-tier API you can call today, and (once weights ship) potentially the most capable model you can self-host.
How does Qwen 3.8 Max compare on benchmarks?
Qwen 3.8 Max's strongest category is general agent work — the multi-step, tool-using tasks that most production AI deployments actually care about. On Alibaba's published benchmark tables it beats GPT-5.6 Sol on five of seven agent benchmarks, including CoWorkBench (74.8 vs 71.5), WorkSpaceBench (67.7 vs 65.6), JobBench (53.4 vs 45.4), and WideSearch (81.9). It also wins on instruction-following (IFBench 82.8 vs 72.7) and several specialized evaluations (HealthBench, PLawBench, PRBench-Finance). But the picture is mixed at the top:
| Benchmark | Qwen 3.8 Max | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|---|
| CoWorkBench | 74.8 | 71.5 | — |
| WorkSpaceBench | 67.7 | 65.6 | — |
| JobBench | 53.4 | 45.4 | 57.4 |
| WideSearch | 81.9 | N/A | — |
| SkillsBench | 70.2 | — | 70.9 |
| IFBench | 82.8 | 72.7 | 63.5 |
| SWE-bench Pro | 67.7 | — | 80.0 |
| FrontierSWE | 73.5 | — | 88.8 |
| DeepSWE 1.1 | 56.6 | — | 70.0 |
| Terminal Bench 2.1 | 86.6 | 88.8 | — |
| PaperBench | 93.0 | 90.5 | — |
| QwenSWEBench | 80.7 | 73.5 | — |
| Humanity's Last Exam | 43.6 | 47.2 | 53.3 |
| MRCR v2 (long-context) | 92.9 | 93.8 | — |
The pattern is clear: Qwen 3.8 Max wins broadly on agent and productivity tasks, but where deep reasoning or the very hardest single-shot coding challenges are the bar, Fable 5 still leads by 10–15 points. Independent verification of these scores (by Artificial Analysis, LMArena, or comparable venues) was not available at the time of writing — every number above is Alibaba's own, so treat them as vendor-reported until third-party tables appear.
For context on why vendor-reported scores warrant caution, see our AI benchmark gaming problem in 2026 breakdown.
How much does Qwen 3.8 Max cost?
The official release API is priced at $2.00 per million input tokens and $6.00 per million output tokens, with cached input at $0.25 per million. That is roughly a third of the rate of Claude Fable 5 and significantly below GPT-5.6 Sol on per-token pricing, and the MoE architecture (only 95 billion of 2.4 trillion parameters active per token) is what makes that price structurally sustainable rather than a promotional discount.
During the July preview, the only access was Alibaba's credit-based Token Plan: Lite at $6 for 2,500 credits per 7 days, and Pro at $68 for 40,000 credits per 7 days with 6–8 concurrent agents. The paid-preview explained much of the early "is it actually open?" confusion. As of the August 3 release, standard per-token API pricing applies via Alibaba Cloud Model Studio and Qwen Cloud, and the preview's credit-only tier is no longer the only option.
For comparison with siblings and the open-weight family:
| Model | Input $/M | Output $/M | Context | Open weights |
|---|---|---|---|---|
| Qwen 3.8 Max | $2.00 | $6.00 | ~1M | promised "next week" |
| Qwen 3.7-Max | $1.25 | $3.75 | 1M | no (closed) |
| Qwen 3.5 Plus | $0.40 | $2.40 | 1M | yes (Apache 2.0) |
| Qwen 3.5 Flash | $0.10 | $0.40 | 1M | yes (Apache 2.0) |
| Kimi K3 | ~$3.00 | ~$15.00 | 1M | yes (Modified MIT, July 27) |
| Claude Fable 5 | higher tier | higher tier | — | no |
For more on how Qwen has historically priced the open-weight family and the self-hosting floor price, see our guide to AI model pricing and the Three Layers Every AI Agent OS Needs writeup, which covers where Qwen fits in a self-hosted stack.
How to access Qwen 3.8 Max right now
There are four practical ways in, in order of friction:
- Qwen Studio (chat.qwen.ai) — free web UI with no signup required; supports text, image, and document input. Best for quick exploration and prompt design. This was the route many early testers used during the preview.
- Alibaba Cloud Model Studio API — the developer endpoint. The model ID is
qwen3.8-max(orqwen3.8-max-previewon some older docs). OpenAI- and Anthropic-compatible, so you can drop it into any existing SDK by changing the base URL and API key. International regions: Singapore (ap-southeast-1, primary), Frankfurt (since March 2026), Hong Kong (since March 2026), US Virginia (no free quota). - Qoder and QoderWork — Alibaba's own agentic coding and productivity platforms, both wired to Qwen 3.8 Max from launch. Useful if you want a turnkey agent harness without wiring MCP or Claude Code yourself.
- Qoder Code (open-source CLI) — the Qwen Code terminal agent at github.com/QwenLM/qwen-code plugs Qwen models into VS Code and the terminal, similar to Claude Code's harness.
Once open weights ship: download from Hugging Face (huggingface.co/Qwen) or ModelScope. For local or self-hosted serving, Qwen documents vLLM and SGLang as the reference runtimes:
# vLLM example once weights are out (check the actual model ID on release)
vllm serve Qwen/Qwen3.8-Max --port 8000 --tensor-parallel-size 8 \
--max-model-len 983040 --reasoning-parser qwen3
Numbers to watch when the model card lands: total parameters 2.4T, active ~95B, documented context window 983,616 tokens (~1M) with a 131,072-token output ceiling per the preview's Codex integration metadata. Those preview specs usually carry over to the release with minor changes.
What is the autonomous coding story?
The single most cited Qwen 3.8 Max result is its 10-day autonomous coding run. The model was pointed at an empty GitHub repository and told to ship a project ("oh-my-cli"). Across 10 days it filed its own issues, wrote code, ran its own tests, and merged its own pull requests without human review. A companion run reproduced a published research paper on data selection for LLM training over 125 hours, then improved on the original paper's AIME24 math result by 2.7 points. A third run entered the Tianchi data-science competition against 526 human teams and beat 458 of them within a 24-hour window.
These are Alibaba's own reports, so apply the usual vendor-claim discount. But the shape of the claim — sustained, self-directed, multi-day work — is exactly the category where Qwen 3.8 Max's benchmark wins sit. That is the headline capability gap that distinguishes it from earlier Qwen Max tiers. If you are building long-horizon coding agents (Claude Code-style harnesses, Hermes Kanban workers, OpenCode flows), the relevant question is not "does it score high on a Q&A test" but "can it stay on-task across hours or days with minimal supervision" — and Qwen 3.8 Max's reported results and benchmark scores both target that exact property. For a concrete blueprint of an agent OS that can swap models like this in and out, see our how to build an agent OS and Hermes + OpenCode free agent team guides.
How does Qwen 3.8 Max fit the 2026 open-weight race?
Three Chinese labs each shipped or previewed a trillion-parameter-class model inside roughly a month:
| Model | Lab | Parameters | Open weights | Standout |
|---|---|---|---|---|
| Kimi K3 | Moonshot AI | 2.8T | yes (Modified MIT, July 27) | #1 on LMArena Frontend Code, beating Fable 5 |
| Qwen 3.8 Max | Alibaba | 2.4T (95B active) | promised "next week" | 10-day autonomous coding; broad agent-task wins |
| GLM 5.2 | Z.AI | 744B | yes (Apache 2.0) | FrontierSWE near Claude Opus 4.8 |
The corporate geometry is unusual: Alibaba holds roughly 36% of Moonshot AI, so Kimi K3 and Qwen 3.8 Max are as much an in-house rivalry as a competitive one. The pattern from the three releases is the real story — there is no longer a single benchmark category where Western frontier labs (OpenAI, Anthropic) hold a comfortable, undisputed lead; the gaps are measured in single digits, not generations, and Chinese labs are doing it while shipping open weights and undercutting on price. Both OpenAI and Anthropic have filed confidentially for IPOs this year, and every release that narrows the frontier gap while publishing open weights and a fraction of the price chips at the story those IPO valuations rest on.
What this means for you
- Small business / solo builder testing the frontier on a budget: Qwen 3.8 Max's $2/$6 per-million-token API is the cheapest credible frontier-tier endpoint running right now. Use the free Qwen Studio web UI to prototype, then move to the API once you know your token budget.
- Developer building long-horizon agents: Qwen 3.8 Max's published strengths are exactly the multi-step agent benchmarks — CoWorkBench, WorkSpaceBench, WideSearch — that map to production agent loops. If your harness supports OpenAI- or Anthropic-compatible APIs, pointing it at Qwen 3.8 Max is a one-config swap.
- Team that needs self-hosting, fine-tuning, or data residency: wait for the open weights. Once they ship under Apache 2.0, a 2.4T-parameter sparse-MoE model with 95B active is large but servable on a serious cluster (probably 8× H200 or equivalent) via vLLM or SGLang. Until then, Qwen 3.5 Plus is the open-weight tier you can actually download today.
- Anyone evaluating frontier models for a purchase: the independently verified number for Qwen 3.8 Max is the architecture scale (2.4T total / 95B active) and the price. Every benchmark score is currently Alibaba's own. Run it against your real workload before committing.
FAQ
Q: Is Qwen 3.8 Max actually open source? A: Not yet as of August 3, 2026. Alibaba says open weights are coming "next week" following the official release. If the release lands, it would be the first Qwen-Max-class model to ship open weights, ending Alibaba's pattern of keeping Max-tier models closed. Until a Hugging Face repo with a license file actually appears, "open" is a promise, not a fact. For context on why this matters for self-hosting, see our benchmark gaming piece.
Q: How does Qwen 3.8 Max compare to Kimi K3? A: Both are Chinese trillion-parameter MoE models released within days of each other in July 2026. Kimi K3 is larger (2.8T vs 2.4T total) and has shipped open weights (Modified MIT) with verifiable third-party benchmarks (93.5% GPQA Diamond, 88.3% Terminal-Bench 2.1). Qwen 3.8 Max has a published benchmark table covering more agent and productivity tasks, a cheaper API ($2/$6 vs roughly $3/$15 per million tokens), and a stronger autonomous-coding narrative. The two labs are also corporate cousins — Alibaba owns ~36% of Moonshot.
Q: How does Qwen 3.8 Max compare to Claude Fable 5? A: Qwen 3.8 Max is competitive on agent and productivity tasks and undercuts Fable 5 significantly on price, but Fable 5 still leads by wide margins on the hardest frontier-coding (SWE-bench Pro 80.0 vs 67.7, FrontierSWE 88.8 vs 73.5) and reasoning (Humanity's Last Exam 53.3 vs 43.6) benchmarks. Qwen 3.8 Max is the open-weight, price-competitive option; Fable 5 is the raw-capability leader on the hardest single-shot tasks.
Q: What does "2.4 trillion parameters" actually mean for cost? A: For a sparse Mixture-of-Experts model, total parameters are not what determines inference cost — active parameters per token are. Qwen 3.8 Max has 2.4T total but only ~95B active per pass, which is what makes the $2/$6 per-million-token price structurally sustainable. A model with more total parameters and fewer active ones can be cheaper to run than a smaller dense model.
Q: Can Qwen 3.8 Max really code autonomously for 10 days? A: Alibaba reports a 10-day run where the model filed, wrote, tested, and merged its own pull requests in a GitHub repo with no human review, plus a 125-hour run that reproduced and then improved on a research paper, improving the original's AIME24 math score by 2.7 points. These are vendor-reported, so apply the usual discount, but the capability the benchmarks and the autonomous runs describe — sustained, self-directed, multi-day work — is consistent and is the model's clearest differentiator vs earlier Qwen tiers.
Q: Is Qwen 3.8 Max free to use? A: The Qwen Studio web UI (chat.qwen.ai) is free with no signup and remains the easiest way to test the model. The developer API on Alibaba Cloud Model Studio is pay-per-token ($2.00/$6.00 per million input/output). New Alibaba Cloud accounts get a 70M+ token free trial (1M tokens per model, valid 90 days, Singapore endpoint only). During the July preview the only developer access was a credit-based Token Plan subscription that has since been superseded by standard per-token pricing.
Discussion
0 comments