Qwen 3.8-Max: Specs, Pricing, Benchmark Status, and How to Access It (2026)

Alibaba previewed Qwen 3.8-Max on July 19, 2026, and officially launched it on August 3 with published specs and API pricing. It is the biggest model release of the summer: 2.4 trillion parameters, multimodal, and a claim that it trails only Anthropic's Claude Fable 5. Here is what is confirmed, what is still a vendor claim, and how to start using it today.

TL;DR

• Launched: August 3, 2026, following a July 19 preview at the World AI Conference in Shanghai
• Size: 2.4 trillion parameters, sparse Mixture-of-Experts, roughly 95B active per token
• Multimodal: text plus visual inputs confirmed. Coverage differs on the full list (video, documents, speech, image generation have all been reported); Alibaba has not published a spec sheet
• Context window: 1M tokens, up to 128k output tokens
• Performance claim: "second only to Fable 5," Alibaba's own words. No benchmark table published yet
• Open weights: landed, in two pieces. Qwen3.8-27B (Apache 2.0) hit Hugging Face August 13-14, and a 2.4T Max-class checkpoint, Qwen3.8-2.4T-A95B, went up around August 12-13 under a custom license. The catch: the 2.4T release is text-only, no vision and no native 1M context
• Access today: standard API at $2 per million input tokens, $6 output, $0.25 cached, plus the Token Plan subscription
• API compatibility: OpenAI spec and Anthropic spec, same as the rest of the Max line
• Running Qwen in production now: Qwen 3.8-Max and Qwen3.8-27B are live on Yotta AI Gateway

What Qwen 3.8-Max Is

Qwen 3.8-Max is the next flagship in Alibaba's Qwen line, previewed two months after Qwen 3.7-Max shipped. Where 3.7-Max was framed entirely around long-horizon agent workloads, 3.8-Max adds a second headline: native multimodality. Text plus visual inputs is confirmed; early coverage also reports video, document, speech, and image generation support, but Alibaba has not published a spec sheet, so treat the full modality list as unsettled.

The stated target workloads are coding, full-stack development, data analysis, and office workflows. That is a direct continuation of the agent positioning the 3.7 line established, now with visual inputs in scope.

The scale claim matters for context: at 2.4 trillion parameters, Qwen 3.8-Max is the second-largest publicly known model, behind Moonshot's Kimi K3 at 2.8 trillion, which launched as an open-weight release the same week. The timing is not a coincidence. The frontier race among Chinese labs is compressing release cycles.

At launch Alibaba disclosed the number that was missing at preview: roughly 95B active parameters per token, about 4 percent of the total. For a sparse MoE model that is the figure that drives serving cost and latency, and it puts 3.8-Max in a lighter serving class per token than Kimi K3's 104B active.

Qwen 3.8 Release Date

Alibaba previewed Qwen 3.8-Max on July 19, 2026, at the World AI Conference in Shanghai, then released it officially on August 3, 2026, with standard API access and published pricing. The preview-only phase is over.

Source: Qwen's official announcement

If you are seeing "Qwen 3.8" and "Qwen 3.8-Max" used interchangeably in coverage, the model shown at WAIC is Qwen 3.8-Max, the flagship tier.

The full split, including the now-released 27B, is in our guide to which Qwen 3.8 version you actually need.

Is Qwen 3.8 Open Source?

Mostly yes, as of mid-August, with one big asterisk on the flagship.

Every Max-tier Qwen model so far has stayed closed. Qwen 3.7-Max is API-only, and the open-weight line continued separately with Qwen 3.6. This time is different: at launch Alibaba committed to releasing open weights within about a week, and not just for the flagship. A smaller Qwen3.8-27B is going open-weight alongside it, and we've broken down the hardware you'll need to run it.

Update: both promises have now landed, within days of each other. Qwen3.8-27B released August 13-14 on Hugging Face under Apache 2.0, a 28B dense model with a vision encoder that runs on a single GPU. And the Max-class drop is real: Qwen3.8-2.4T-A95B appeared on Hugging Face around August 12-13, under a custom qwen3.8-max license rather than Apache. Read the fine print before celebrating, though. The open 2.4T checkpoint is text-only, with no vision input and no native 1M context, which is exactly what the most heated Hugging Face discussion threads are about. And at 2.4 trillion parameters, even aggressive 1-bit quants sit above 400GB, so running it is a datacenter project, not a workstation one. For the full multimodal Max, the API remains the only route. If you want open weights you can actually deploy, Qwen3.8-27B is the practical pick from this family, with GLM 5.2 and Kimi K3 as the alternatives.

Qwen 3.8-Max Benchmarks

Alibaba launched Qwen 3.8-Max without an official benchmark table, and that is worth saying plainly.

Alibaba's exact words: Qwen 3.8 is "one of the most powerful models available today, comparable to leading frontier AI models, second only to Fable 5," referring to Anthropic's Claude Fable 5. That ranking rests on internal evaluations. For the Max API model, that is still true: no official benchmark table. The open 2.4T checkpoint is a different story. Its model card ships with benchmark tables, GPQA Diamond 92.6 among them, the first official numbers attached to this tier. Treat card numbers as claims until the independent indexes pick it up.

For calibration, the verified scores of its predecessor: Qwen 3.7-Max posted 92.4 on GPQA Diamond, 80.4 on SWE-bench Verified, and 69.7 on Terminal-Bench 2.0, competitive with Claude Opus 4.6 on most of the agentic suite. If 3.8-Max improves on that baseline while adding multimodality, the ranking claim is plausible. Plausible is not verified.

The right move for production teams: wait for independent numbers, then run your own workload against it.

We track every verified data point as it lands in our Qwen 3.8 benchmarks breakdown.

How to Access Qwen 3.8-Max

Qwen 3.8-Max is generally available through Alibaba's standard API at $2 per million input tokens, $6 per million output, and $0.25 for cached input. The Token Plan subscription and the Qoder platforms remain the subscription-style alternative.

The model speaks both the OpenAI API spec and the Anthropic API spec, consistent with the rest of the Max line, so existing client code should port with minimal changes.

Qwen 3.8-Max itself is now live on Yotta AI Gateway: one API key, OpenAI-compatible and Anthropic-compatible endpoints, routed alongside Claude, DeepSeek, GLM, and the rest of the catalog. Qwen3.8-27B is on the Gateway too if the flagship is more model than the task needs. We broke down the full deployment decision in how to run Qwen 3.7 in production, and the same logic will apply to 3.8 as access expands.

We also keep a live breakdown of Qwen 3.8 API access options, including Token Plan preview pricing, updated as availability changes.

Qwen 3.8-Max vs Qwen 3.7-Max

What actually changed, based on what has been disclosed:

The practical read: 3.7-Max is the Qwen you can build on today. 3.8-Max is the one to evaluate the moment real benchmarks appear. If cost is the deciding factor between frontier models, our Qwen 3.7 Max vs Claude Opus 4.6 pricing breakdown shows how the current generation compares.

Qwen 3.8-Max vs Kimi K3

The two biggest model announcements of July landed days apart, and they represent opposite bets.

Moonshot's Kimi K3 shipped first at 2.8 trillion parameters, the largest open-weight model ever announced, with weights released July 27.

Qwen 3.8-Max previewed days later at 2.4 trillion parameters then launched August 3 as a managed API at $2 in / $6 out per million tokens, undercutting K3's $3 in / $15 out. The open-weight side has since landed as the text-only Qwen3.8-2.4T-A95B.

For production teams, the choice is the same one the GLM 5.2 vs Qwen 3.7-Max matchup posed: open weights give you deployment control, fine-tuning rights, and GPU-level cost management; a closed frontier API gives you the vendor's best model with zero infrastructure lift.

K3 already has real third-party results while Qwen 3.8's numbers remain vendor claims, we break down everything verified so far in our full Kimi K3 guide.

We break the two models down in full in our Qwen 3.8 vs Kimi K3 comparison.

Frequently Asked Questions

When did Qwen 3.8 release? Alibaba previewed Qwen 3.8-Max on July 19, 2026, at WAIC in Shanghai, and released it officially on August 3, 2026, with standard API access and published pricing.

Is Qwen 3.8 open source? Partly. Qwen3.8-27B is on Hugging Face under Apache 2.0 and runs on a single GPU. A 2.4T Max-class checkpoint, Qwen3.8-2.4T-A95B, is also up under a custom license, but it is text-only and drops the vision and native 1M context of the Max API. The full multimodal Max remains API-only.

How big is Qwen 3.8-Max? 2.4 trillion parameters in a sparse Mixture-of-Experts architecture, with roughly 95B active per token. Active parameters are the number that drives serving cost.

How much does Qwen 3.8 cost? The standard API is $2 per million input tokens, $6 per million output, and $0.25 for cached input. The Token Plan subscription is the alternative for individual use.

Is Qwen 3.8-Max better than Qwen 3.7-Max? Alibaba says it is, claiming it trails only Claude Fable 5 overall. No official benchmark table has been published, so the claim cannot be verified yet. Qwen 3.7-Max's published scores remain the verified baseline for the family.

Can I use Qwen 3.8-Max through an API? Yes. Standard API access is live at published pricing, the model supports both OpenAI and Anthropic API specs, and Qwen 3.8-Max is also live on Yotta AI Gateway. The Token Plan and Qoder platforms remain the subscription route.

Is Qwen 3.8 bigger than Kimi K3? No. Kimi K3 is 2.8 trillion total parameters with 104B active; Qwen 3.8-Max is 2.4 trillion with roughly 95B active. Total parameter count says little about quality or serving cost on its own, especially for sparse MoE models where active parameters matter more.

How can I run Qwen models in production today? Qwen 3.8-Max and Qwen3.8-27B are both live on Yotta AI Gateway with OpenAI-compatible and Anthropic-compatible endpoints, alongside 3.7-Max and 3.6 Plus. Open-weight Qwen models can also be self-hosted on Yotta GPU Pods.

Bottom Line

All three open questions from launch week are now closed: pricing published, specs disclosed, and the open weights actually dropped, in a stripped, text-only form. What remains is independent benchmarking of the flagship.

Qwen you can ship on is 3.8-Max itself, live on Yotta AI Gateway today. Start there, and this post will be updated as 3.8 access expands. The next generation is already peeking out: Qwen 3.8-Flash-Next is the first look at Qwen 4. For the date and what Alibaba has actually confirmed, see Qwen 4: release date and what's known.