Verdict: Qwen 3.8 Max is Alibaba's most capable open-weights model to date — a 2.4-trillion-parameter sparse mixture-of-experts (MoE) architecture with 95 billion active parameters, a 1-million-token context window, and native multimodal understanding of text, images, video, and documents. Released August 3, 2026, with open weights promised on Hugging Face and ModelScope the week of August 10, it is the first time Alibaba has open-licensed a Max-tier model. On vendor-reported benchmarks it leads on PaperBench (93.0%), MathVision (95.2%), and OSWorld-Verified (86.1%), but trails Anthropic's Claude Opus 5 on Toolathlon-Verified (72.5% vs 80.6%) and SWE-bench Pro (67.7% vs Claude's 80.3%). For builders priced out of Claude Fable 5 or GPT-5.6 Sol, Qwen 3.8 Max offers a credible frontier-class alternative at roughly a third the per-token cost — but the scores are vendor-reported, and independent replication had not landed at publish time.
Last verified: 2026-08-04 · Best open-weight frontier: Qwen 3.8 Max · Best cheap open-weight coding: DeepSeek V4 Flash ($0.14/M) · Best price-to-context: Kimi K3 ($3/$15, 1M context, open weights) · Pricing/limits change often — re-check before committing.
What is Qwen 3.8 Max?
Qwen 3.8 Max is the flagship of Alibaba's Qwen 3.8 generation, built on the Qwen 3.5 architecture scaled to 2.4 trillion total parameters. It uses a sparse Mixture-of-Experts (MoE) design: only 95 billion parameters — about 4% of the total — activate for any single token, which is what keeps inference tractable on something this large. It is Alibaba's first multimodal model above 1 trillion parameters, accepting text, images, video, and documents as input (Qwen release blog, Aug 3 2026).
The model was first previewed as qwen3.8-max-preview on July 19, 2026 at the World AI Conference in Shanghai via Alibaba's official X account (@Alibaba_Qwen, July 19 2026), and the production model went live August 3, 2026 on QwenCloud, Qoder, and QoderWork. Open weights are promised "next week" from the August 3 release — placing the Hugging Face and ModelScope drop around August 10, 2026. Critically for builders, the model speaks both OpenAI and Anthropic API protocols natively, so existing coding-agent harnesses and tooling can target it without a rewrite (Qwen release blog, Aug 3 2026).
How does Qwen 3.8 Max compare to Kimi K3 and GLM-5.2?
For most builders choosing an open-weight frontier model in 2026, the decision comes down to three Chinese releases that landed within weeks of each other. Each has a genuinely different shape.
| Model | Total params | Active params | Context | Open weights | API cost (in/out per 1M) | License | Source |
|---|---|---|---|---|---|---|---|
| Qwen 3.8 Max | 2.4T | 95B | 1M | Promised ~Aug 10, 2026 | $2 / $6 | TBD (announced) | Qwen blog, OpenRouter |
| Kimi K3 | 2.8T | 16 of 896 experts | 1M | Released July 27, 2026 | $3 / $15 | Modified MIT | Moonshot platform docs, Kimi pricing |
| GLM-5.2 | 744B | not disclosed | 1M | Already open | $1.40 / $4.40 | MIT | Z.ai release |
| DeepSeek V4 Flash | not disclosed | not disclosed | 1M | Already open | ~$0.14 / $0.28 | DeepSeek License | Morph |
The takeaway: Qwen 3.8 Max sits between Kimi K3 (larger, pricier, weights already shipped) and GLM-5.2 (much smaller, much cheaper, MIT-licensed and ready today). On raw parameter scale K3 is bigger (2.8T vs 2.4T); on price-per-token Qwen undercuts K3 by roughly a third ($2/$6 vs $3/$15 via OpenRouter); on license clarity GLM-5.2 wins outright because its MIT terms are already on the page while Qwen and K3's open-weight license details are announced but not yet published.
For builders who want the deeper architectural comparison, Kimi K4's roadmap and the chip bottleneck that may decide it covers how the supply chain shapes this race.
What are the benchmark scores — and can you trust them?
Here is the part every Qwen article glosses over: all scores below are vendor-reported. Qwen published them; independent evaluators (Artificial Analysis, LMArena) had not replicated Qwen 3.8 Max numbers at time of writing. Treat them as directional, not settled.
| Benchmark | Qwen 3.8 Max | Best peer (vendor-reported) | Source |
|---|---|---|---|
| Terminal-Bench 2.1 (coding agent) | 86.6% | Qwen 3.8 Max (best of compared set) | Qwen blog |
| PaperBench (paper reproduction) | 93.0% | Qwen 3.8 Max | Qwen blog |
| OSWorld-Verified (OS agent) | 86.1% | Qwen 3.8 Max | Qwen blog |
| MathVision (multimodal) | 95.2% | Qwen 3.8 Max | Qwen blog |
| SWE-bench Pro (real-world coding) | 67.7% | Claude Mythos 5 — 80.3% | Qwen blog |
| Toolathlon-Verified (tool use) | 72.5% | Claude Opus 5 — 80.6% | Qwen blog |
| LongBench v2 (long-context) | 66.3% | Qwen 3.8 Max | Qwen blog |
| MMMU-Pro (multimodal reasoning) | 82.3% | GPT-5.4 Pro — 94.0% | Qwen blog |
The honest read: Qwen 3.8 Max wins on the agentic-OS and paper-reproduction lanes, is competitive on long-context, but is meaningfully behind Claude Opus 5 on tool-orchestration and SWE-bench Pro — and behind GPT-5.4 on multimodal reasoning. Benchmarks that the same vendor sells and scores are a marketing artifact until a third party reproduces them; The AI Benchmark Gaming Problem in 2026 covers why this matters more each cycle.
That said, there is one number that is not a benchmark and is genuinely interesting: Qwen's release blog reports that an oh-my-cli GitHub repository ran fully autonomously — the model writing its own commits, PRs, and issue triage — for roughly 16 days ending July 30, producing 265 commits, 127 PRs, and 151 issues with no human engineer in the loop (Qwen blog, Aug 3 2026). That is a duration claim that is easy to verify independently by reading the repository, unlike a leaderboard score that requires Qwen's own harness to reproduce.
How much does Qwen 3.8 Max cost?
Qwen 3.8 Max API pricing is $2 per million input tokens and $6 per million output tokens via OpenRouter and other third-party providers, with a 1,000,000-token context window and a 131,072-token maximum output (OpenRouter, Aug 2026).
If you prefer Alibaba's own surfaces, the entry point is the QwenCloud Token Plan — a credit-based monthly subscription rather than a clean per-token rate. Tiers (approximate USD from CNY):
| Tier | Monthly | 7-day credit quota | 5-hour quota |
|---|---|---|---|
| Lite | ~$6 (39 CNY) | 2,500 credits | 700 credits |
| Standard | ~$20 (139 CNY) | 10,000 credits | 3,000 credits |
| Pro | ~$70 (499 CNY) | 40,000 credits | 12,000 credits |
Token Plan works with any tool that speaks OpenAI or Anthropic protocol — Qwen Code, Cline, Claude Code, Cursor, OpenCode, and Codex are all officially supported (QwenCloud Token Plan). During the preview window, credit consumption for the Max model ran at 10% of the standard rate, with an additional 80%-off applied between 22:00 and 08:00 UTC+8 — a promotional price, not the long-term rate.
If your goal is to run a frontier coding agent with zero budget rather than minimum budget, How to Run DeepSeek V4 Flash for Free in 2026 covers five no-cost paths to a model that scores 82.7 on Terminal-Bench 2.1 at $0.14 per million tokens when you do want to pay.
Can you run Qwen 3.8 Max locally?
Realistically, no — not the full Max checkpoint, even once the weights ship. At 2.4 trillion parameters in FP16 you need roughly 4.8 TB of VRAM just to load the weights, which is well beyond anything but a multi-node H200 or B200 cluster. The MoE active parameter count (95B) helps inference cost on a hosted route, but it does not help you load the model — you still need the full 2.4T on disk and in memory.
The companion 27B Qwen 3.8 model announced alongside Max is the practical self-hosting target. Alibaba said both the Max and 27B variants will go open weights the same week (Qwen blog, Aug 3 2026). For a perspective on what a 276B-parameter open-weights model that "matches its 975B sibling" actually delivers, see Inkling-Small (2026): The Open-Weights Model That Matches Its 975B Sibling at a Quarter the Size — the same distillation tradeoff applies here at a different scale.
Why is Alibaba open-sourcing a Max-tier model now?
Three forces are converging in mid-2026:
The Chinese open-weights wave. DeepSeek V4 Flash, Moonshot's Kimi K3 (2.8T, MIT, weights shipped July 27), and Zhipu's GLM-5.2 (744B, MIT, already open) all landed within weeks of each other. China's labs are competing to put frontier-grade weights on Hugging Face at a pace US labs have not matched. Qwen 3.8 Max is Alibaba's answer: the first Max-tier model the company has ever open-licensed (@Alibaba_Qwen, July 19 2026).
The "Pacing the Frontier" petition. On July 28, 2026, more than 1,100 employees from OpenAI, Anthropic, Google, Meta, Microsoft, Mistral, and Thinking Machines signed a letter asking the US government to back an international "brake" on automated AI development. OpenAI and Anthropic formally endorsed the petition on July 29 (Axios, July 30 2026; Fortune, July 29 2026). Their stated concern is recursive self-improvement — models automating more of the research that builds stronger successors — accelerating capability beyond human control. If US labs slow down, the open-weights supply from China becomes structurally more important, not less.
The ai2027.com prediction, revisited. Published April 3, 2025, the AI 2027 scenario forecasts a superhuman coder by 2027 and offers two endings: "slowdown" and "race." The race ending — in which no lab can afford to stop because the competitor is two months behind — is the one the authors consider more plausible, and the one recent release cadence is consistent with (ai-2027.com; Timelines Forecast, April 2025).
You do not have to buy the doomer framing to notice the pattern: a frontier-class open-weights model now ships roughly every two to four weeks. The "AI fatigue" some developers describe — too many model drops to meaningfully evaluate any single one — is a real occupational side-effect of that cadence, not a marketing trope.
What this means for you
| If you are... | Do this |
|---|---|
| A solo builder priced out of Claude/GPT | Run Qwen 3.8 Max through OpenRouter at $2/$6 once the API is live; if cost is zero, DeepSeek V4 Flash is still the free-tier benchmark. |
| A team running coding agents in production | Wait for independent benchmarks before routing production traffic. The 16-day autonomous oh-my-cli run is the claim most worth verifying yourself on a real repo. |
| Self-hosting on a single GPU | Target the 27B companion, not Max. The 95B active count is misleading — you need the full 2.4T checkpoint loaded to serve it. |
| Choosing a long-term open-weights bet | GLM-5.2 already has MIT terms and runs cheaper; Kimi K3 has open weights today at $3/$15. Qwen 3.8 Max's license terms are still TBD — confirm them on Hugging Face before committing. |
| Worried about model-release pace outpacing your ability to evaluate | You are not wrong. The practical fix is a stable eval harness you can re-run on each release, so you spend an hour on a benchmark and not a week chasing every drop. |
FAQ
Q: Is Qwen 3.8 Max actually open source? A: Alibaba has announced open weights for the week of August 10, 2026, on Hugging Face and ModelScope, but has not published the license terms yet. "Open weights" is the company's phrasing. Until the model card lands, "open source" is a promise, not a fact. Compare Kimi K3's Modified MIT terms (already on the page) and GLM-5.2's MIT terms (live since release).
Q: Does Qwen 3.8 Max beat Claude Opus 5? A: On vendor-reported agentic-OS and paper-reproduction benchmarks, yes or competitive. On tool-orchestration (Toolathlon-Verified 72.5% vs Opus 5's 80.6%) and real-world coding (SWE-bench Pro 67.7% vs Claude Mythos 5's 80.3%), no. Independent replication had not landed at the time of writing.
Q: How is Qwen 3.8 Max different from Qwen 3.8-Max-Preview? A: Preview was the July 19, 2026 hosted early-access version on Alibaba's Token Plan, Qoder, and QoderWork. The August 3 general-availability model — the one going open weights — is the production successor, available through OpenRouter, ofox, and other third-party routes at $2/$6 per million tokens.
Q: What hardware do I need to run Qwen 3.8 Max locally? A: Realistically you cannot run the full Max checkpoint on a single machine. The 2.4T-parameter model needs roughly 4.8 TB of VRAM in FP16 — multi-node H200 or B200 territory. The companion 27B variant, also going open weights the same week, is the practical self-host target.
Q: Is Qwen 3.8 Max multimodal? A: Yes. It accepts text, images, video, and documents as input and is Alibaba's first multimodal model above 1 trillion parameters. It does not generate images or video as output.
Q: Should I switch from Claude Code or OpenAI Codex to Qwen 3.8 Max? A: Not as a blanket recommendation. Qwen speaks both OpenAI and Anthropic protocols, so pointing an existing harness at it is a low-friction experiment — but bench your actual workload, not the vendor's chosen benchmarks. The AI Benchmark Gaming Problem in 2026 covers why a 5-point benchmark lead rarely survives contact with a real repo.
Discussion
0 comments