Alibaba Launches Qwen3.8-Max: 2.4T MoE, PaperBench SOTA at 93.0, All Coding Evals Run Through Claude Code
Alibaba’s Qwen team launched Qwen3.8-Max on August 3, its largest and most capable model to date. The model carries 2.4 trillion total parameters with 95 billion active per query — a mixture-of-experts architecture built on the Qwen3.5 base. It is multimodal across text, images, and video, accepts up to 1 million tokens of context, and is priced at $2 per million input tokens, $6 per million output. Open weights arrive on Hugging Face and ModelScope next week.
The headline benchmark: PaperBench at 93.0, the highest score in Alibaba’s published comparison and 2.5 points ahead of GPT-5.6 Sol (90.5) and 4.2 points ahead of Fable 5 (88.8). Everything else in the comparison table needs more context before it can be taken at face value.
Benchmark Table
All numbers below come from Alibaba’s own internal runs. Artificial Analysis had not published independent results as of the launch.
| Benchmark | Qwen3.8-Max | Opus 4.8 | Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal Bench 2.1 | 86.6 | 84.6 | 84.6 | 88.8 |
| SWE-bench Pro | 67.7 | 69.2 | 80.0 | 64.6 |
| DeepSWE 1.1 | 56.6 | 59.0 | 70.0 | 73.0 |
| FrontierSWE | 73.5 | 70.0 | 88.8 | — |
| PaperBench | 93.0 | 80.3 | 88.8 | 90.5 |
| GPQA Diamond | 92.6 | 92.0 | 92.6 | 94.1 |
| IFBench | 82.8 | 62.2 | 63.5 | 72.7 |
| HLE | 43.6 | 45.7 | 53.3 | 47.2 |
| OSWorld-Verified | 86.1 | 83.4 | 85.0 | 83.2 |
The shape: Qwen leads on PaperBench, IFBench, and OSWorld-Verified. It trails Fable 5 by 12.3 points on SWE-bench Pro and by 15.3 points on FrontierSWE. On the deepest software-engineering rows, Alibaba’s own table puts it mid-pack or lower. Its Terminal Bench score (86.6) beats both Anthropic flagships but trails Sol at 88.8.
The Claude Code Harness Detail
Alibaba ran most coding evaluations using Claude Code, pointed at Qwen3.8-Max through its Anthropic-compatible API. This is disclosed in the launch materials and matches the standard practice of evaluating models inside widely-used tooling. It also means Qwen’s coding scores are partially shaped by Anthropic’s scaffold design, and that the harness advantage is the same one applied to Opus 4.8 and Fable 5 in the same table — so the comparisons are internally consistent, but none of them would necessarily hold under a different harness.
Arena Scores (Independent)
These are the only third-party numbers as of August 3:
- Frontend Code Arena: #4 at 1,668 Elo ±18 (1,563 votes, preliminary)
- Text Arena: #5 at 1,496 Elo
- Vision Arena: #2 globally (behind Claude Fable 5 only)
For context, Kimi K3 sits at 1,676 in the same Frontend Code leaderboard with significantly more votes (4,366). The Qwen score will tighten with additional votes, but the early position is within overlap range of K3 — not clearly ahead or behind.
What This Model Is
Qwen3.8-Max is the first Max-class Qwen model with open weights. Every prior Qwen-Max release stayed closed-API. Alibaba framed the launch around long-horizon autonomous work: a 16-day coding run that produced 265 commits and 127 pull requests in an open-sourced agent framework (oh-my-cli), a 125-hour research paper reproduction that improved on the source paper’s AIME24 result by 2.7 points, and a simulated year-long e-commerce operation that finished 38% ahead of GLM 5.2.
These cases are not benchmarks — they are Alibaba’s own test environments with no external verification. They are framed as demonstrations of agentic capability over multi-day horizons, a direction that Moonshot (Kimi K3) and Anthropic (Fable 5 agent systems) are pursuing in parallel.
Pricing and Positioning
At $2/$6, Qwen3.8-Max is 7.5x cheaper than Kimi K3 ($3/$15) and 25x cheaper than Fable 5 ($15/$75 standard). Against Sol at $5/$30, it is 2.5x cheaper on input. The quality gap for agentic coding work — measured by the available SWE-bench Pro and DeepSWE rows — favors US frontier models, but the cost gap is large enough to anchor a different deployment tier.
What Comes Next
Artificial Analysis was not yet scoring Qwen3.8-Max as of August 3. The AA Intelligence Index comparison — Kimi K3 sits at 57, and Qwen3.8 has no published score — is the main outstanding question. Independent SWE-bench Verified results will clarify where it sits against the models it claims to match on Terminal Bench. The weights, once released, will also determine whether the model’s performance holds up under community harnesses.