Qwen3.8 Max is worth a controlled pilot when you need long-horizon reasoning, coding, document or image understanding, and tool use. It is not a free pass to faster work: the current evidence points to a capable but slow and verbose route, with access and pricing changing by snapshot and provider.
The decision anchor is Artificial Analysis's dated Qwen3.8 Max (0902) snapshot: Intelligence Index 45, 37.2 output tokens per second, $5.41 per Intelligence Index task, and 190M output tokens across its v4.3.2 evaluations. That is the practical trade-off—capability purchased with a long reasoning trace—not a universal latency or invoice. (Artificial Analysis, Alibaba model catalogue)
Key takeaways
Qwen3.8 Max is the newer Qwen route released August 3, 2026; the current independent page identifies a September
0902snapshot.Alibaba's live catalogue lists it for text generation and image/video understanding. That catalogue presence does not prove every account, provider or Tabbit picker exposes it.
The 3.8 family is described by BenchLM as open-weight, 2.4T total / 95B active, with a 1M context; hosted API and self-hosting are separate decisions.
Artificial Analysis reports Index 45, 37.2 tok/s, $2/$6 per million tokens and $5.41 per index task. BenchLM reports a different 73.2/100 aggregate and 40 tok/s. Do not merge the two leaderboards.
Qwen Studio free access, QwenCloud API, Token Plan subscriptions, third-party providers and Tabbit Browser have different billing and policy boundaries.
No Tabbit Qwen3.8 Max task or screenshot was completed for this draft. Start with the Qwen3.8 Max model resource.
What Qwen3.8 Max actually is
Qwen3.8 Max is Alibaba/Qwen's flagship route for reasoning, coding, knowledge work and agent workflows, with multimodal input listed in the current catalogue and independent records. The hosted model ID, open-weight checkpoint, Qwen Studio chat route and a provider alias are not interchangeable. The Qwen3.8 Max prompts and reviews preserve the model-specific resource trail.
| Question | Current snapshot | Boundary |
|---|---|---|
| Model ID / release | qwen3.8-max; BenchLM says Aug. 3, 2026 | The 0902 alias is a later dated route; pin it in logs. |
| Family | Qwen3.8-2.4T-A95B, roughly 95B active according to BenchLM/source notes | Parameter claim is source-specific; hosted and open-weight deployment differ. |
| Context | BenchLM tracks 1M; AA tracks 984K for 0902 | Client, route and output ceiling can be smaller. |
| Input | Text; Alibaba catalogue also lists image/video understanding | Multimodal harness and provider support must be checked. |
| Reasoning | Reasoning model in AA/BenchLM records | Effort, tools and fallback affect cost and latency. |
| API price | AA snapshot $2 input / $6 output per million; Alibaba search exposes another $1.65/$4.951 lane | Exact official table was not readable; never turn provider price into official price. |
| Access | Alibaba Model Studio catalogue, Qwen Studio/QwenCloud and providers | Account, region, quota, subscription and policy differ. |
Qwen3.8 Max versus Qwen3.7 Max
The comparison is a dated product boundary, not a promise that every task improves.
| Dimension | Qwen3.7 Max | Qwen3.8 Max | Decision |
|---|---|---|---|
| Documentation anchor | May 20, 2026 snapshot | Aug. 3 release; 0902 independent snapshot | Pin both dates before comparing. |
| Input boundary | Text-only API in the checked 3.7 page | Text/image input in current catalogue and AA record | Test the actual provider multimodal route. |
| Context | 1M / 131,072 max output in checked 3.7 API | BenchLM 1M; AA 984K; output max not sourced here | Do not copy 3.7 output limits into 3.8. |
| Model shape | Hosted Max route | Open-weight 2.4T/95B-active family plus hosted SKU | Self-host cost and hosted token cost differ. |
| Evidence anchor | 201 tok/s AA comparison in older article | AA 37.2 tok/s and $5.41/index task for 0902 | Speed and cost are snapshot-specific. |
Alibaba's live catalogue currently lists qwen3.8-max for text generation and image/video understanding. This is the cleanest current access fact. Qwen's release page was reopened but rendered only its site shell, so detailed release claims remain source notes until the page can be read again. The Qwen3.7 Max overview is context, not a substitute for 3.8 evidence.
Benchmarks: read the denominator
Artificial Analysis v4.3.2 reports 45/200, rank 17/200, 37.2 tok/s, $5.41 per index task and 190M output tokens. Its index combines ten named evaluations and a specific reasoning/provider snapshot. BenchLM reports 73.2/100, rank 8/230, 40 tok/s, 53.3 seconds first token and strong multimodal/grounded coverage, while explicitly saying it has no comparable first-party API rate. These are different catalogues with different denominators and evidence states. The agentic reasoning guide is more useful than a pasted “best model” label.
How to get it without mixing products
Alibaba Model Studio/DashScope: confirm region, model ID, thinking setting, cache, tool schema, quota and current pricing table.
Qwen Studio: a chat or coding view can offer a free or promotional route; it does not prove API entitlement or unlimited use.
QwenCloud or another provider: record provider ID, fallback, cache policy, data policy and price. Provider rates are not Alibaba's list price.
Token Plan/subscription: treat seat or monthly usage as separate from pay-as-you-go API billing.
Tabbit Browser: its subscription/download is separate from all Qwen routes; this article did not verify the live model picker.
Unknown risks and scenario self-check
Alias drift:
0902and a future alias can be different snapshots; pin the ID and date.Slow, verbose runs: AA's 37.2 tok/s and 190M evaluation output tokens are a dated warning, not an SLA.
Multimodal harness: catalogue support does not define resolution, tools, provider or evaluator conditions.
Access and policy: the Reddit MCP thread disputes “free/unlimited” claims and reports account-policy friction. Verify current terms; do not automate around a website.
Open weight versus hosted: hardware, license, quantization and provider routing create different total costs.
| Your situation | First move | Do not infer |
|---|---|---|
| Long-context code or research | Bounded, sourced task with a stop rule | 1M context does not make every answer cheaper. |
| Multimodal document work | Same images and rubric through your route | Catalogue support is not identical provider support. |
| Budget-sensitive API | Compare one cached and one uncached run | AA/BenchLM task cost is not your invoice. |
| Qwen Studio/MCP user | Disposable folder and diff review | “Free” does not mean unlimited or policy-safe. |
| Browser workflow | Read the agentic browser guide and confirm permissions | Qwen capability does not grant authentication. |
What users report
In the coding discussion, users call it one of the best, complain that plans consume weekly limits after a few prompts, report two-to-eight-minute thinking steps, and describe strong web-game and Rust-editor generation. These are conflicting personal observations with no shared harness. In the Qwen Studio + MCP thread, the author says the workflow works but is slower than Codex/Claude Code; commenters dispute unlimited access and discuss account suspension. X exposed no stable quote, and YouTube exposed test titles/chapters only, so neither is hard evidence.
A practical next step
Pick one reversible task: a small repository change with tests, a sourced document comparison, or a fixed image-understanding fixture. Record model ID, dated snapshot, provider, reasoning setting, modality, tokens, latency, tools, retries, policy prompts and human corrections. Run it beside Qwen3.7 Max or your current model. Keep Qwen3.8 Max only when the accepted result justifies the extra time, tokens and review.
For browser work, use browser automation and Tabbit Browser as the product layer. This article did not verify an account-level Tabbit route. Compare broader AI browser choices before depending on a model selector.
Tabbit BrowserVerdict
Qwen3.8 Max is a serious pilot candidate for long-horizon reasoning and multimodal work. The useful headline is not “2.4T parameters”: it is that a dated independent snapshot pairs Index 45 with slow output, high evaluation token use and $5.41 per task. Compare accepted work, not a family label. Keep Alibaba API pricing, Qwen chat/subscription access, providers and Tabbit separate, and require a reversible test before production use.
Sources
FAQ
What is Qwen3.8 Max?
Qwen3.8 Max is Alibaba's Qwen flagship route for reasoning, coding, knowledge work and multimodal input. The current Alibaba catalogue lists qwen3.8-max for text generation and image/video understanding.
What changed from Qwen3.7 Max?
The 3.8 route is positioned as a newer Aug. 3 release with a 2.4T-total/95B-active open-weight family, multimodal input and a later 0902 snapshot. The 3.7 Max article is a separate May snapshot and text-only API boundary.
How much does Qwen3.8 Max cost?
Artificial Analysis currently shows a dated 0902 route at $2 input/$6 output per million tokens. Alibaba's search result also exposes a separate $1.65/$4.951 pricing lane, but the target pricing page was not readable in this pass; confirm the region and route before budgeting.
Where can I access Qwen3.8 Max?
Alibaba Model Studio currently lists qwen3.8-max. Qwen Studio, QwenCloud, providers and Token Plan subscriptions are separate routes with their own limits and terms.
Is Qwen3.8 Max good for coding?
It merits a controlled coding and multimodal pilot, but independent data describes it as slow and verbose and community reports are mixed. Use a fixed repository, tools, stop rule and acceptance test.
Can I use Qwen3.8 Max in Tabbit?
This overview did not run an account-level Tabbit test and makes no availability claim. Check the live picker and run a small reversible task first.