Qwen3.8 Max
Alibaba's hosted flagship: a 2.4T MoE that sits just under the Western frontier and above every other open-lab model on Arena WebDev. Priced like a mid-tier US model, not like a Chinese one. Strong at coding and MCP tool use.
Key facts
- Score
- 78 / 100, rank #11Evals90Price42People62Fit85
- Price
- $2 in · $6 out per 1M tokens (Qwen3.8 Max (0902) - Artificial Analysis)
- Blended price
- $3 per 1M tokens (3 input : 1 output)
- Context window
- 984K
- Open weights
- No
- Released
- Aug 3, 2026
- Best for
- Agentic coding, Tool/MCP-heavy agents, Multilingual production chat
Qwen3.8 Max costs $2 per million input tokens and $6 per million output tokens on Alibaba's list price (Qwen3.8 Max (0902) - Artificial Analysis, read Oct 11, 2026).
Independent evals
| Benchmark | Result | Who ran it | Read |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 45 | Qwen3.8 Max (0902) - Artificial Analysis | Oct 10, 2026 |
| Arena Text score | 1483 (rank 22) | Text Arena Leaderboard | Oct 10, 2026 |
| Arena WebDev score (qwen3.8-max-0902) | 1674 (rank 10) | WebDev Arena Leaderboard | Oct 10, 2026 |
| Scale MCP Atlas (Qwen3.8-2.4T-A95B open-weight variant, xHigh) | 84.50 | MCP Atlas Leaderboard - Scale Labs | Oct 10, 2026 |
What practitioners say
The launch thread (1,124 points, 612 comments) was dominated by people already running the smaller Qwen 3.6/3.8 line locally and cancelling Claude subscriptions; the Max itself drew praise as the first Qwen-Max-class model with open weights. Skeptics (Aurornis) say Qwen output 'has to be discarded for anything other than really easy tasks', and several note that at current electricity prices self-hosting is pricier than DeepSeek's cache rate.
Caveats
- At $2/$6 it costs 4-5x GLM-5.3 or MiMo-V2.6-Pro for a similar AA score.
- The hosted Max (0902) is proprietary; the open-weight sibling is Qwen3.8-2.4T-A95B, which scores lower on AA (40) and needs serious hardware.
- Slow: ~35 tok/s on AA's measurement.
Questions people ask
How much does Qwen3.8 Max cost?
Qwen3.8 Max costs $2 per million input tokens and $6 per million output tokens on Alibaba's list price (Qwen3.8 Max (0902) - Artificial Analysis, read Oct 11, 2026).
How good is Qwen3.8 Max?
Qwen3.8 Max ranks #11 of 26 on the Models at Work leaderboard with a score of 78 out of 100 (October 2026 edition). Alibaba's hosted flagship: a 2.4T MoE that sits just under the Western frontier and above every other open-lab model on Arena WebDev. Priced like a mid-tier US model, not like a Chinese one. Strong at coding and MCP tool use.
What is Qwen3.8 Max best for?
Qwen3.8 Max is best for agentic coding, tool/mcp-heavy agents, multilingual production chat.
What are Qwen3.8 Max's benchmark scores?
Artificial Analysis Intelligence Index: 45 (Qwen3.8 Max (0902) - Artificial Analysis); Arena Text score: 1483 (rank 22) (Text Arena Leaderboard); Arena WebDev score (qwen3.8-max-0902): 1674 (rank 10) (WebDev Arena Leaderboard); Scale MCP Atlas (Qwen3.8-2.4T-A95B open-weight variant, xHigh): 84.50 (MCP Atlas Leaderboard - Scale Labs).
What is Qwen3.8 Max's context window?
Qwen3.8 Max has a 984K token context window according to Alibaba.
Is Qwen3.8 Max open weights?
No. Qwen3.8 Max is a proprietary model available through Alibaba's API and partner clouds.
Compare Qwen3.8 Max
Methodology: Score is 0 to 100 and computed, not typed: 55% independent evals (Artificial Analysis, Arena, Scale SEAL, Epoch, each scaled against its natural floor and the best score in this table, then averaged), 15% blended price on a fixed log scale ($0.05 per million is 100, $60 is 0), 15% practitioner sentiment from Signals and community threads, 15% operational fit (context window, open weights). A missing component drops out and the rest are reweighted. A model with no independent eval yet is provisional and ranks below every measured one. Within three points is a tie. Vendor-published figures are labelled and never counted.