Sign in Subscribe
news

Qwen 3.8

Sebastien Dubois

Sebastien Dubois

— 5 min read

Canonical version: Qwen 3.8.

Qwen3.8-Max is a 2.4 trillion parameter multimodal LLM from Alibaba's Qwen team. Announced 19 July 2026 as a preview, generally released on 3 August 2026 under the title "Qwen3.8-Max: A New Bar for Coding and Cowork".

The July version of this note was a list of missing information. Most of it now exists.

What shipped

  • 2.4T total parameters, 95B active. The active count was the conspicuous gap in July; it is now published. 95B active on a 2.4T sparse MoE is a ~26:1 sparsity ratio
  • Built on the Qwen 3.5 architectural foundation, not a new line
  • Open weights promised for the week of 10 August. This would be the first time Alibaba open-sources a Max-class model, after keeping 3.6-Max and 3.7-Max API-only
  • 1M token context: 991K max input, 131K max output, 262K max reasoning tokens
  • Multimodal in (text, images, video), text out
  • Function calling, context caching, structured outputs, web search, code interpreter, web extraction
  • Available through QwenCloud

Pricing

Per 1M tokens
Input $2.00
Output $6.00
Implicit cache read $0.25
Explicit cache write $2.50
Explicit cache read $0.17

Rate limits: 2M TPM, 15K RPM. On Hacker News this was measured at roughly 40% of Kimi K3's price for the same model class, which is the number that will matter most in practice.

Benchmarks

Alibaba published a full table against Claude Opus 4.8, Claude Fable 5, GPT-5.6 Sol and Qwen3.7-Max. Selected rows, Qwen3.8-Max last:

Benchmark Opus 4.8 Fable 5 GPT-5.6 Sol Qwen3.7-Max Qwen3.8-Max
Terminal Bench 2.1 84.6 84.6 88.8 74.5 86.6
SWE-bench Pro 69.2 80.0 64.6 60.6 67.7
DeepSWE 1.1 59.0 70.0 73.0 21.6 56.6
FrontierSWE 70.0 88.8 — 40.7 73.5
PaperBench 80.3 88.8 90.5 64.8 93.0
CoWorkBench 72.3 75.9 71.5 64.6 74.8
WideSearch 72.9 81.2 — 75.2 81.9
GPQA Diamond 92.0 92.6 94.1 92.4 92.6
HLE 45.7 53.3 47.2 41.4 43.6
IFBench 62.2 63.5 72.7 79.1 82.8
HealthBench 52.4 — 55.3 54.5 60.2
MRCR v2 256K 83.2 — 93.8 86.7 92.9

The honest read: frontier-adjacent, not frontier-leading, at roughly a third of the price. It takes PaperBench, WideSearch, IFBench, HealthBench and the legal/finance rows. It loses clearly on SWE-bench Pro, DeepSWE and HLE, where Fable 5 is well ahead. The jump over Qwen3.7-Max is the largest story in the table, particularly DeepSWE at 21.6 → 56.6.

The three coding demonstrations

These are the substance of the announcement, and they're more interesting than the benchmark table.

16 days of autonomous coding. The model built oh-my-cli from an empty folder: a self-evolving harness where requirements become issues, agents claim them through a state machine, and code, tests, previews and logs feed back in. By 30 July, roughly 16 days of unattended operation had produced 265 commits, 127 PRs and 151 issues. The full trace is public. Alibaba's own term for the setup is Loop Engineering (see Loop Engineering).

Reproducing a paper, then beating it. Given only a paper and GPUs, no starter code, it spent ~125 hours writing ~7,600 lines, took 1,100+ actions and ran 33 rounds of GPU training. First it reproduced all six findings (~37 hours). Then it ran a hypothesis → code → train → analyse loop, inventing 18 improvement ideas over four rounds, and finished with a method that beats the paper's own by +2.7 points on AIME24.

526 human teams, 24 hours. Entered into a live Tianchi competition, it fine-tuned and ensembled BERT, MacBERT and RoBERTa for text plus Qwen2.5-VL-7B and Chinese-CLIP for images, fused them by weighted voting, and climbed from 0.60 to 0.853 across 45 submissions, beating 458 of 526 human teams.

Harness generalisation

Worth noting for anyone building on top: Alibaba explicitly scaled its RL environments across harnesses, and reports comparable performance on QwenWork, Claude Code, Codex CLI, OpenClaw and Hermes Agent. Most of the benchmark rows above were in fact measured through the Claude Code harness.

A model tuned to work well inside someone else's harness, rather than only its own, is a different competitive posture than we've seen from most labs.

How it looked in July (kept as history)

The preview announcement landed two days after Moonshot AI shipped Kimi K3 at 2.8T with weights that actually appeared on Hugging Face. At that point Qwen 3.8 had no weights, no model card, no license, no published pricing, no benchmarks and no independent evaluation. This note called it "a sentence rather than a repository".

That judgment was right for the moment and is now largely settled: the numbers and the pricing arrived on 3 August. The weights have not. That remains the open item.

Why it matters

The strategy shift held. Chinese labs built their reputation on small, fast, cheap and good enough. Qwen 3.8 and Kimi K3 are the opposite bet, and the follow-through here (a real price, a real benchmark table, a dated weights commitment) makes it harder to read as pure flag-planting.

Commoditising the model to sell the surface. If Alibaba Cloud is the best place to build on Qwen, the weights are customer acquisition, not the product. Pricing at ~40% of the nearest competitor is consistent with that.

Cheap frontier-adjacent capability changes agent economics. As one Hacker News comment put it, API cost is ceasing to be the bottleneck for agentic workflows. A model that is 90% as good at 33% of the price is the better default for long-horizon agent loops where token burn, not peak reasoning, decides the bill.

The counter-pressure still stands. Open weights concentrated in one region erodes the incentive to train models elsewhere. Open weights from several geopolitical regions is the outcome to want, and it is still not the outcome trending.

Caveats

  • Every benchmark is Alibaba's. No independent evaluation yet
  • The Terminal Bench comparison is not apples to apples. Qwen3.8-Max was measured with Claude Code at avg@10 with a 5-hour timeout; competitors are "best published score across harnesses" pulled from Artificial Analysis and OpenAI. Different harness, different sampling
  • Five of the benchmarks are in-house (QwenSWEBench, QwenQoderBench, QwenReactBench, QwenSVGBench, plus the Qoder UX bench). Unfalsifiable by outsiders
  • A footnote states Fable 5 results "may involve fallbacks", which is doing quiet work in a table where Fable 5 is the main rival
  • The weights are still a promise. Second dated commitment; the first slipped
  • Token-hungriness of this model generation is unmeasured here. Effective cost may sit well above the headline
  • Qwen 3.8 27B is reported as arriving the same week, which is the release most people will actually run locally

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe