Quick answer. Qwen3.8-Max is Alibaba's 2.4T-parameter MoE model (95B active), released August 3, 2026, with a 262K context extensible to 1M and a permissive custom licence. It is elite at algorithmic coding and tool use — LMArena Code #3 — but its headline agentic benchmarks do not replicate under neutral harnesses, and it cannot realistically be run locally.
Qwen3.8-Max is the current flagship of Alibaba's Qwen family, teased on July 19, 2026 as a preview and released on August 3. It is a large release by any measure, and the launch numbers were strong enough to top Hacker News.
They also do not entirely survive independent testing. This guide covers what Qwen 3.8 is, what it is genuinely excellent at, and where the launch claims fall apart — because the gap between the two is unusually instructive.
What is Qwen3.8-Max?
| Spec | Qwen3.8-Max |
|---|---|
| Parameters | 2.4T total / 95B active (MoE) |
| Experts | 512 (10 routed + 1 shared) |
| Layers | 92 |
| Architecture | Hybrid Gated-DeltaNet / full attention, built on the Qwen3.5 stack |
| Context | 262,144 native, extensible to 1,010,000 |
| Released | Preview July 19, 2026 · GA August 3, 2026 |
| Licence | Custom qwen3.8-max — not Apache 2.0 |
What is the licence, really?
Better than the "custom licence" label suggests, and this matters more to most teams than any benchmark.
The terms are close to verbatim MIT, with two thresholds bolted on:
- Attribution is required above 100M monthly active users or $20M monthly revenue.
- A separate licence is required only if you run a Model-as-a-Service or "AI Work Assistant" business above $50M in twelve months. The clause explicitly names Alibaba's own Qoder and QwenWork products, so its target is competing coding-assistant vendors.
Internal use is carved out entirely. For the overwhelming majority of teams — anyone building an internal tool, a product below those thresholds, or a service that is not a competing AI assistant — Qwen3.8-Max is free and commercially usable.
It is still accurate to call this open weights rather than open source. The training data is not published and the licence is not OSI-approved. But the practical freedom is real.
Can you run Qwen 3.8 locally?
No, and it is worth being blunt because the Qwen family's reputation is built on local use.
| Format | Size | Realistic hardware |
|---|---|---|
| BF16 | 4.89 TB (213 shards) | Datacentre cluster |
| FP8 | 2.50 TB (213 shards) | Datacentre cluster |
| 1-bit quant (Unsloth) | 397 GB | ≥450 GB RAM; ~20 tok/s on B200s |
The smallest quantization anyone has produced still needs more memory than a well-specced server, and it runs at roughly 20 tokens per second on datacentre GPUs. There is no verified tokens-per-second figure for consumer hardware because nobody can run it on consumer hardware.
The model most people actually want — Qwen3.8-27B — has not been released. Its Hugging Face repository still returns a 401 as of August 13, 2026. If you are here for a local Qwen upgrade, the honest answer is to wait, or look at Muse Glimmer or Qwen 3.7 in the meantime.
Do the launch benchmarks hold up?
Two of the headline claims do not, and the way they fail is genuinely worth understanding.
The "best agentic model" claim expired within hours
Qwen3.8-Max did briefly top Artificial Analysis's Agentic Index on August 6 — by 0.1 points (55.4 to 55.3), on a sub-index that at the time contained only two benchmarks, which the two models split cleanly between them.
AA published a methodology update the same day and confirmed the demotion on the record. As of August 13, Qwen3.8-Max ranks 3rd, behind Claude Opus 5 (59.17) and Grok 4.6 (58.68). Any article still repeating the original claim is a week and two positions out of date.
The Terminal-Bench result reverses under a neutral harness
This is the more serious one. Alibaba claimed 86.6% on Terminal-Bench, against 84.6% for Claude Opus 4.8 — a 2-point win.
Alibaba's own footnote states that the Opus 4.8 comparison figure was taken from Artificial Analysis's Terminus 2 harness. AA also ran Qwen on Terminus 2:
| Harness | Qwen3.8-Max | Claude Opus 4.8 | Result |
|---|---|---|---|
| Alibaba's own setup | 86.6% | 84.6% (borrowed) | Qwen +2.0 |
| AA Terminus 2 (both models) | 81.3% | 84.6% | Qwen −3.3 |
| AA, using Alibaba's Claude Code harness | 79.4% | — | |
| Vals AI | 67.4% | — |
Measured the same way, a 2-point win becomes a 3.3-point loss. The comparison mixed one model's score from a favourable in-house setup with another's from a neutral one.
The mechanism appears to be a 5-hour timeout disclosed in Alibaba's own footnotes. The official Terminal-Bench leaderboard's rules state that submissions may not modify timeouts or resources — and Qwen3.8-Max was never submitted to that board, whose all-time best score is 83.8%.
An important nuance: Qwen's static benchmarks replicate fine. Independent GPQA came in at 92.6% against a claimed 92.7%. It is specifically the harness-dependent agentic numbers that inflate — which is exactly the pattern we documented with Meta's Muse Spark and with xAI. Long-horizon agentic benchmarks are simply much easier to flatter than one-shot ones.
It is not a coding leader on its own numbers
Read Alibaba's own model card carefully and Qwen3.8-Max leads 1 of 12 coding rows — PaperBench. It loses all four of Alibaba's own in-house coding benchmarks to Claude Fable 5.
So what is Qwen 3.8 genuinely good at?
Quite a lot, and this is the part the benchmark controversy obscures.
- Algorithmic coding — genuinely elite. LMArena Code rank #3 (Elo 1668.6), and #3 of 58 on IOI-style competitive programming. If your work is algorithm-shaped, it is excellent.
- Overall LMArena: #8 of 391.
- Tool use: #1 of 20 on τ³-Banking. Genuinely best in class.
- Price: $2 / $6 per million tokens for a frontier-scale model.
- Context: 1M extensible, matching the frontier.
Where it falls down
- Agentic software work. AA's Coding Agent Index puts it #15 of 55, and agentic coding is its worst LiveBench category. SWE-bench Verified is 85.6% against Claude Opus 5's 97.0%.
- Hallucination. AA-Omniscience scores it 3.4 against Claude Opus 5's 37.1 — a very large gap on factual reliability. This is the finding least covered elsewhere and the one most likely to bite in production.
- Cost per completed task. Despite $2/$6 tokens it is the second-most-expensive of 20 models per completed task, because it burns far more tokens getting there. Cheap tokens, expensive jobs — the same trap we flagged with Grok 4.6.
- Not locally runnable, and the 27B version is unreleased.
Should you use Qwen 3.8?
Yes, if your work is algorithmic — competitive programming, data structures, optimisation, self-contained problems. Top-3 on LMArena Code at $2/$6 is a strong offer.
Yes, if you need heavy tool calling. #1 on τ³-Banking is not a marginal lead.
Yes, if the licence matters and you are below the thresholds — near-MIT terms on a frontier-scale model is rare.
No, if you want a long-horizon coding agent. #15 of 55 on the Coding Agent Index and an 11-point SWE-bench gap to Opus 5 are decisive.
No, if factual reliability is critical. The AA-Omniscience gap is too large to design around.
No, if you wanted to run it locally. Wait for Qwen3.8-27B.
The honest summary
Qwen3.8-Max is a specialist wearing a generalist's launch chart. It is genuinely excellent at algorithmic coding and tool use, priced well, and licensed generously. It is mediocre at agentic software engineering, unreliable on facts, and expensive per completed job.
The wider lesson is the useful one: Qwen's static benchmarks replicated and its agentic benchmarks did not. That is now a pattern across three vendors in two weeks. When a launch chart leads with a long-horizon agentic number, assume the harness is doing some of the work until someone neutral reproduces it.
FAQ
What is Qwen3.8-Max?
Alibaba's flagship model, released August 3, 2026. A 2.4-trillion-parameter mixture-of-experts model with 95B active parameters, a 262K native context extensible to about 1M, and a permissive custom licence.
Is Qwen 3.8 open source?
It is open weights, not open source. The custom qwen3.8-max licence is close to MIT, requiring attribution above 100M MAU or $20M monthly revenue, and a separate licence only for Model-as-a-Service or AI-assistant businesses above $50M a year. Internal use is exempt.
Can I run Qwen 3.8 locally?
Realistically no. BF16 weights are 4.89 TB and FP8 is 2.50 TB. The smallest community quantization is a 397 GB 1-bit build needing at least 450 GB of RAM. The 27B version people are waiting for has not been released.
Is Qwen 3.8 the best agentic model?
No. It briefly led Artificial Analysis's Agentic Index on August 6 by 0.1 points on a two-benchmark sub-index, and AA revised its methodology the same day. It now ranks 3rd, behind Claude Opus 5 and Grok 4.6.
Why do Qwen 3.8's Terminal-Bench scores differ so much?
Harness differences. Alibaba reported 86.6% from its own setup while borrowing its Claude comparison figure from Artificial Analysis. Run on AA's neutral Terminus 2 harness, Qwen scores 81.3% — turning a 2-point win into a 3.3-point loss. Vals AI measured 67.4%.
Is Qwen 3.8 good for coding?
For algorithmic coding, yes — LMArena Code #3 and IOI #3 of 58. For agentic software engineering, no: AA's Coding Agent Index ranks it #15 of 55, and SWE-bench Verified is 85.6% against Claude Opus 5's 97.0%.
How much does Qwen 3.8 cost?
$2 per million input tokens and $6 per million output. But it is the second-most-expensive of 20 models measured per completed task, because it consumes considerably more tokens per job.
When is Qwen3.8-27B coming?
No date has been announced. Its Hugging Face repository still returns a 401 as of August 13, 2026.