Qwen3.8-Max is Alibaba’s August 3, 2026 flagship and, at 2.4 trillion parameters (95B active), the biggest model the Qwen team has published with benchmarks. The pitch: “second only to Fable 5” — and on agentic and research benchmarks, the numbers back it up. On deep software engineering? Not so much.
The real story is the combination: frontier-competitive scores at $2/$6 per million tokens — roughly a third of Claude Opus 5 and a quarter of GPT-5.6 Sol — with open weights (and a 27B dense sibling) released within two weeks of launch — Alibaba delivered on the open-weight pledge, making Qwen3.8-Max one of the most capable downloadable models. Qwen3.7-Max is its immediate predecessor, now outpaced across the board.
Architecture
| Spec | Value |
|---|---|
| Total parameters | 2.4T |
| Active parameters | ~95B per token (~4% of network) |
| Architecture | Sparse MoE transformer (based on Qwen 3.5) |
| Context window | 1M tokens |
| Input modalities | Text, image, video |
| Output modalities | Text |
| API model ID | qwen3.8-max |
| Reasoning effort | low / medium / high |
Pricing
| Rate / 1M tokens | |
|---|---|
| Input | $2.00 |
| Output | $6.00 |
Compatible with both OpenAI and Anthropic API protocols on QwenCloud.
Benchmarks
All numbers are vendor-reported from Alibaba’s August 3 launch post. Independent validation is still limited — treat these as directional, not settled.
| Benchmark | Qwen 3.8-Max | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|---|---|
| TerminalBench 2.1 | 86.6 | 84.6 | 88.8 | 84.6 |
| SWE-bench Pro | 67.7 | 80.0 | 64.6 | 69.2 |
| PaperBench | 93.0 | 88.8 | 90.5 | 80.3 |
| GPQA Diamond | 92.6 | 92.6 | 94.1 | 92.0 |
| OSWorld-Verified | 86.1 | 85.0 | 83.2 | — |
| IFBench | 82.8 | 63.5 | 72.7 | 62.2 |
| HLE (Humanity’s Last Exam) | 43.6 | 53.3 | 47.2 | 45.7 |
The pattern: Qwen 3.8-Max dominates agentic/research rows (OSWorld, PaperBench, IFBench) but trails Fable 5 badly on core software engineering (SWE-bench Pro gap of 12.3 points, FrontierSWE gap of 15.3 points). “Second only to Fable 5” is accurate on some axes, misleading on others.
The open weights
Alibaba’s launch post promised “the first open-weight model at Max scale” the following week, and it shipped:
- Qwen3.8-Max (
Qwen/Qwen3.8-2.4T-A95B) — the full 2.4T model, released on August 12, 2026 in BF16 and FP8 under a customqwen3.8-maxlicense. - Qwen3.8-27B — the dense variant for self-hosting on more accessible hardware, released on August 14, 2026 under Apache 2.0.
The open checkpoint is the post-trained text model with 92 layers (a Gated DeltaNet and Gated Attention hybrid), 512 experts with 10 routed plus 1 shared active, and a 262,144-token native context extendable to about 1.01M. The hosted Qwen3.8-Max is “the official version based on” it with vision input, non-thinking support, a 1M-token default context and built-in tools.
Alibaba has since followed with Qwen3.8-Flash-Next, a much smaller open model that previews the Qwen4 architecture.
How it compares
- vs Claude Fable 5: Fable wins deep coding by double digits. Qwen wins agentic computer use and research tasks. Fable costs $10/$50 — 5× more on input, 8× on output.
- vs GPT-5.6 Sol: Sol leads TerminalBench (88.8 vs 86.6) and GPQA Diamond (94.1 vs 92.6). Qwen leads OSWorld and PaperBench. Sol costs $5/$30 — 2.5×/5× more.
- vs Claude Opus 5: Opus 5 sits at $5/$25 with broader agentic tooling. Qwen matches or exceeds on most benchmarks at 40% of the price.
- vs Kimi K3: Kimi K3 is 2.8T (larger) and also open-weight; both are now downloadable. Qwen 3.8-Max claimed higher benchmark scores at launch, but Xiaomi’s MiMo-V2.6-Pro later reported surpassing both on the Artificial Analysis index.
- vs Qwen3.7-Max: The predecessor. Qwen 3.8-Max improves TerminalBench from 74.5 to 86.6, PaperBench from 64.8 to 93.0, SWE-bench Pro from 60.6 to 67.7. A generational leap.
Limitations
- SWE-bench gap is real. 67.7 on SWE-bench Pro means Qwen 3.8-Max is not the model to pick for deep, autonomous software engineering — Fable 5 at 80.0 remains the clear choice for that job.
- Vendor benchmarks only. The full table was published alongside the launch with no independent replication yet.
- Open weights are huge. The 2.4T checkpoint is a multi-node deployment; most self-hosters will use the 27B variant or Qwen3.8-Flash-Next instead.
- Multimodal scope unclear. “Text + visual inputs confirmed” — but whether video, documents and speech are all natively supported or vary by mode is not fully specified.
For the full AI model landscape, see the 2026 LLM guide.
Best for
- Long-horizon autonomous agents and computer-use tasks
- End-to-end research reproduction and scientific workflows
- Cost-effective frontier inference at $2/$6 vs $5/$25+ proprietary alternatives
- Downstream fine-tuning and self-hosting on the released open weights
Pros & cons
Strengths
- Top scores on agentic benchmarks — OSWorld-Verified 86.1 (ahead of Fable 5's 85.0 and GPT-5.6 Sol's 83.2)
- PaperBench 93.0 — highest reported score for end-to-end research reproduction
- Aggressive pricing at $2/$6 — roughly 1/3 of Opus 5 and 1/4 of GPT-5.6 Sol
- Open weights released August 12 (plus an Apache-2.0 27B variant on August 14) — the first open-weight Max-class model from Alibaba
Limitations
- SWE-bench Pro 67.7 — trails Fable 5 by 12+ points and Opus 4.8 by ~1.5 points on deep coding
- Benchmarks are vendor-reported — limited independent validation so far
- The 2.4T open weights need a large multi-GPU cluster to serve, and use a custom Qwen3.8-Max license rather than Apache 2.0
- Multimodal input scope (video, documents, speech) not fully specified by Alibaba
Frequently asked questions
When was Qwen 3.8-Max released?
Alibaba previewed Qwen 3.8-Max on July 19, 2026 at the World AI Conference in Shanghai. The full launch with published specs, benchmarks and API access came on August 3, 2026.
How big is Qwen 3.8-Max?
2.4 trillion total parameters in a sparse Mixture-of-Experts configuration, with approximately 95 billion parameters active per token (~4% of the network firing per query). Built on the Qwen 3.5 architecture.
Is Qwen 3.8-Max open source?
Yes — the weights were released after launch. Qwen3.8-2.4T-A95B (BF16, 2.4T parameters, 95B active) went up on Hugging Face on August 12, 2026 under a custom `qwen3.8-max` license, and the dense Qwen3.8-27B followed on August 14 under Apache 2.0. The open checkpoint has a 262K native context (extendable to about 1.01M); the hosted Qwen3.8-Max adds vision input, non-thinking mode, a 1M default context and built-in tools.
How does Qwen 3.8-Max compare to Fable 5?
Qwen 3.8-Max leads on OSWorld-Verified (86.1 vs 85.0), PaperBench (93.0 vs 88.8) and TerminalBench (86.6 vs 84.6). But Fable 5 crushes it on SWE-bench Pro (80.0 vs 67.7) and FrontierSWE (88.8 vs 73.5) — a 12–15 point gap on exactly the deep coding benchmarks that matter most for software engineering agents.
What is reasoning_effort?
Qwen 3.8-Max supports three reasoning effort levels: high (complex tasks demanding thorough analysis), medium (balancing accuracy and cost), and low (efficient reasoning optimized for speed). This maps to the same concept as Anthropic's effort tiers.