News / Model Watch · Open-Weights Release
Qwen3.8 Open Weights (2026): Alibaba Drops the 2.4T Max Checkpoints — and an Apache-2.0 27B You Can Run at Home
When we briefed Qwen3.8-Max on Aug 8, the open-weights promise was a "coming soon" footnote. It's here now: the Qwen Hugging Face org carries Qwen3.8-2.4T-A95B (plus an FP8 variant) — the first open-weight Max-class model from any lab — and Qwen3.8-27B, a dense, Apache-2.0, image-and-video-capable 27B that runs from a 17GB file. This brief has the timeline, the license fine print, the API pricing, and a field-tested local-run recipe.
Qwen3.8's weights are open as of Aug 12–17, 2026: Qwen3.8-2.4T-A95B (+ FP8) is the first Max-class open checkpoint, and Qwen3.8-27B ships under Apache 2.0 with native image/video understanding and 262K context (extensible to 1M).
The 27B is the one you can actually self-host — Simon Willison ran it on a 128GB MacBook at 15–30 tok/s and called it "by far the best pelican SVG I've been able to generate with a model that runs on a local machine."
Hosted qwen3.8-max API pricing per launch coverage: $2 / $6 / $0.25 per 1M tokens (input / output / cached).
From announcement to open weights: the two-week timeline
| When | What happened · source |
|---|---|
| Early Aug (Aug 3 per CCTV) | Qwen3.8-Max announced and served via API — "most capable model to date," open weights promised. Our Aug 8 brief covered the launch and the rumor-mixup with Qwen2.5-VL. |
| Week of Aug 12 | The official blog post "Qwen3.8-Max: A New Bar for Coding and Cowork" marks the first open-weight Max-class release; the 2.4T-A95B and FP8 checkpoints appear on the Qwen HF org in the days following (community timelines: eigent.ai, codersera). |
| Aug 14–16 | Qwen3.8-27B lands under Apache 2.0. Simon Willison posts a hands-on (Aug 16) that hits the front page of Hacker News at 798 points; Artificial Analysis lists the 27B at 52 on its Intelligence Index. |
| Aug 17 (Mon) | CNBC frames it against Meta: "Alibaba answers Meta's AI challenge with new laptop-ready model." Hugging Face tells CNBC that Qwen-based derivatives hit 151,448 — 2.6× Meta's footprint. |
Dates follow the sources cited per row; where community timelines disagree by a day (Aug 12 vs 14 for the exact checkpoint hours), we say so instead of flattening it.
What shipped, per the official model cards
- Qwen3.8-2.4T-A95B (the Max-class drop). 2.4 trillion total parameters, built on the Qwen3.5 architecture foundation. Official benchmark tables put it alongside Opus 4.8, Claude Fable 5, and GPT-5.6 Sol: Terminal Bench 2.1 at 86.6 (GPT-5.6 Sol max: 88.8), PaperBench at 93.0 (the top score in the table), GPQA Diamond at 92.6, Agents' Last Exam at 53.6. Artifacts are documented as compatible with vLLM, SGLang, and TokenSpeed.
- Qwen3.8-27B (the one you can run). Dense 27B, native vision-language — images and videos — with flexible thinking control, 262,144-token native context extensible to 1,000,000. Alibaba's launch materials claim it "outperforms Qwen3.7-Plus overall"; community consensus (per codersera's Qwen guide) treats it as the new default local Qwen over Qwen3.6-27B.
- The hosted-vs-weights split matters. Per the official card,
qwen3.8-max(the hosted version) is built on the 2.4T checkpoint and adds vision input, non-thinking mode, 1M context by default, and official built-in tools. If you self-host the raw checkpoint, you're serving the base — not the full hosted feature set. - FP8 variants of the big checkpoint are published alongside, for datacenter inference teams that want the smaller memory footprint.
What didn't ship — read the fine print
- The 2.4T license is not (reported) Apache. The 27B is Apache 2.0; community guides flag the 2.4T repos as shipping under a custom Qwen license instead. As of Aug 21 we could not extract the full license text from the repo page — read the model card before commercial self-hosting, especially if you serve it externally.
- Hosted 27B: no date, no price. The official card says a Qwen Cloud hosted 27B is "coming soon" with 1M context by default and built-in tools — nothing more specific at publication.
- "2.4T open" ≠ "2.4T in your GPU box." Community guides put the FP8 checkpoint around 2.5TB — multi-node territory. The open-sourcing is strategically huge and practically relevant to inference platforms, not to your workstation.
- Pricing attribution. The $2 / $6 / $0.25 per-1M figures come from consistent launch coverage (latent.space, eigent.ai, codersera, kie.ai against the Qwen Cloud listing) — not from a pricing page we captured directly. Verify on Model Studio before locking budgets.
Qwen3.8 API pricing vs the Chinese-model shelf (2026)
At a reported $2.00 input / $6.00 output per 1M tokens (and $0.25 cached input), qwen3.8-max sits in the same price band as its frontier peers — comfortably under Western flagships, above DeepSeek's off-peak floor.
| Model (2026) | Input $/1M | Output $/1M | Open weights |
|---|---|---|---|
| qwen3.8-max | $2.00 | $6.00 | 2.4T-A95B + FP8 on HF (custom license reported) |
| GLM-5.3 (Z.ai) | $1.40 | $4.40 | promised ~Aug 28 (our brief) |
| deepseek-v4-flash (off-peak) | $0.22 | $0.66 | MIT (V4 family) |
| Kimi K3 | $3.00 | $15.00 | 2.8T, modified license (our brief) |
Comparison rows reuse prices already verified in our earlier briefs (GLM-5.3, DeepSeek, Kimi K3); the qwen3.8-max row is launch-coverage-reported — see fine print above.
Run it at home: Simon Willison's field-tested recipe
- The setup. The 17GB
Q4_K_MGGUF in LM Studio, tested on a 128GB MacBook Pro (M5 Max) and an NVIDIA DGX Spark — roughly 15–30 tokens/sec, and about 72% faster with llama.cpp's draft-mtp speculative decoding. - Fix the context default first. 262,144 native tokens, but LM Studio's 8,192 default "was exhausted by reasoning alone." Raise it before you judge the model.
- Kill the overthinking. Default reasoning effort is xhigh: his pelican-SVG prompt ran 21 minutes and burned 22,276 reasoning tokens to emit 3,223 output tokens. "It's a great model, but wow that default setting is a bad place to start" — use low or no reasoning for everyday tasks.
- The verdict. "By far the best pelican SVG I've been able to generate with a model that runs on a local machine" — and on vision: his bounding-box tests against pelican photos were "a very good match." The catch is dense-model speed: memory bandwidth-bound, so it won't feel like a small MoE. As he put it: "The fact that a 17GB file can do all of this stuff on my home machines is a miracle."
Which Qwen3.8 should you route to?
| Your workload | Route to | Why |
|---|---|---|
| Local dev, privacy-sensitive, offline | Qwen3.8-27B self-hosted | Apache 2.0, 17GB Q4, vision+video input; tune thinking down from xhigh |
| Hardest coding/agent work, hosted | qwen3.8-max API | Full hosted feature set (vision input, non-thinking, 1M default, built-in tools) at $2/$6 |
| Cheapest solid text/agent loop | deepseek-v4-flash | $0.22/$0.66 off-peak — see our Flash family brief; watch the new peak/off-peak windows |
| Max-class open weights, self-inferenced | 2.4T-A95B / FP8 | For inference platforms with multi-node capacity; vLLM/SGLang/TokenSpeed documented |
How the story landed: overseas and at home
- Hacker News made it a moment. Willison's review sits at 798 points; the Artificial Analysis listing (52 on the Intelligence Index) adds a 380-point thread. The recurring themes: best-in-class local capability, and the xhigh-default quirk.
- Western framing is Meta-centric. CNBC's angle is competition with Meta's new Muse Glimmer laptop models; Hugging Face's 151,448 Qwen-derivatives stat (2.6× Meta) does the arguing. WIRED and Axios, covering the same open-weights wave this month via GLM-5.3, now treat Chinese open models as the default frontier of open AI.
- Chinese coverage puts it in a one-week trifecta. 北京商报's Aug 16 roundup — "一周三更" — lines up DeepSeek's repricing, Alibaba's open-sourcing, and Zhipu's post-training play in one week, with OpenRouter data showing 6 of the global top-10 models by usage that week being Chinese.
Get Qwen and the whole Chinese shelf through one OpenAI-compatible endpoint
ChinaModelAPI is an independent relay that brings Chinese frontier models — Qwen, DeepSeek, GLM, Kimi — to builders worldwide behind a single OpenAI-compatible API, with USDT/USD1 payments and no subscription. Drops like this one are exactly what the platform is built for: day-one model availability, transparent per-token pricing, and one integration for the whole shelf.
ChinaModelAPI is an independent relay with no official relationship to Alibaba or the Qwen team. Routed model IDs are verified live before launch — join the waitlist to be notified when Qwen3.8 routing goes live.
Primary sources
- Qwen official blog — "Qwen3.8-Max: A New Bar for Coding and Cowork" (first open-weight Max-class release)
- Hugging Face — Qwen/Qwen3.8-2.4T-A95B model card (benchmarks, vLLM/SGLang compatibility, hosted-Max feature split)
- Hugging Face — Qwen/Qwen3.8-27B model card (dense VLM, 262K→1M context, hosted version "coming soon")
- Hugging Face — Qwen organization (460 models; FP8 variant listed)
- Simon Willison — "Qwen 3.8 27B is excellent, but it defaults to overthinking things" (Aug 16, 2026)
- CNBC — "Alibaba answers Meta's AI challenge with new laptop-ready model" (Aug 17, 2026; HF 151,448 derivatives stat)
- Artificial Analysis — Qwen3.8-27B listing, Intelligence Index 52
- Hacker News — Willison review thread (798 points)
- eigent.ai — Qwen3.8-Max open-weight summary (pricing $2/$6/$0.25 per latent.space)
- codersera — Qwen generation guide (27B Apache 2.0, supersedes Qwen3.6-27B; 2.4T license caveat)
- kie.ai — 27B launch breakdown (Alibaba "outperforms Qwen3.7-Plus" claim, HF card specs)
- 北京商报/东方财富 — "大模型一周三更" (Aug 16; OpenRouter top-10 stat)
FAQ (2026)
Are the weights really open?
Yes — 2.4T-A95B (+FP8) and the 27B are on the Qwen HF org. The 27B is Apache 2.0; the 2.4T is reported to carry its own Qwen license terms, so check the card before commercial self-hosting.
Can I run the 2.4T myself?
Not on workstation hardware — the FP8 checkpoint is around 2.5TB per community guides. It's for inference platforms; everyone else uses the hosted API or the 27B.
Hardware for the 27B?
17GB Q4_K_M GGUF; Willison used a 128GB MacBook Pro (M5 Max) and a DGX Spark at 15–30 tok/s, +72% with draft-mtp. Raise the context default — 8,192 gets eaten by reasoning alone.
Images and video input?
Yes — the official 27B card describes native vision-language understanding of images and videos, plus flexible thinking control. Rare at this size class.
What does the API cost?
Launch coverage consistently reports $2 / $6 / $0.25 per 1M (input / output / cached) for qwen3.8-max on Qwen Cloud. Verify the live Model Studio page before budgeting.
Hosted 27B when?
"Coming soon" per the official card — 1M context default plus built-in tools. No date or price as of Aug 21, 2026.
Why the overthinking complaints?
xhigh reasoning is the default: 22,276 reasoning tokens and 21 minutes for one SVG. Willison's fix: start at low or no reasoning and only escalate when needed.
vs Kimi K3 / GLM-5.3?
K3 opened 2.8T weights (modified license) July 27; Qwen3.8 is first with a Max-class checkpoint plus an Apache dense sibling; GLM-5.3 weights are due ~Aug 28. All three tracked on our Model Watch.