Wednesday, November 4

Latest in AI

Ask Handy AI 2 free questions a day
Model Drop

Qwen3.8-Max

Qwen3.8-Max
2 min read
Qwen3.8-Max

Qwen3.8-Max, which Alibaba first made public as a preview on July 19, 2026 and opened to API customers on August 3, is a 2.4-trillion-parameter multimodal flagship that Alibaba ranks second only to Claude Fable 5.

  • Model: Qwen3.8-Max (qwen3.8-max; previewed as Qwen3.8-Max-Preview)

  • Maker: Alibaba

  • Ship date: July 19, 2026 (preview; API August 3, 2026)

  • Pricing: $2 / $6 per M tokens (API from August 3)

  • Context: 1M tokens

  • Where you can use it:

  • When to reach for it:

    • Long-horizon coding and office agent work

    • Big-context jobs mixing text, images, video

    • Skip it for the hardest repo-scale coding

  • How it compares:

    • Alibaba ranks it just behind Claude Fable 5

    • Cheaper per token than Kimi K3

  • What makes it different:

    • 2.4T parameters, 95B active per token

    • First Max-class Qwen with open weights

  • Early reactions: Early coverage was skeptical, since the preview shipped with no benchmarks or model card behind the "second only to Fable 5" claim.

  • Read more: Qwen3.8-Max: A New Bar for Coding and Cowork, model card

The benchmarks

Alibaba published these with the August 3 API release (the July 19 preview had none): its own runs, mostly in the Claude Code harness, against Claude Opus 4.8, Claude Fable 5 (results "may involve fallbacks") and GPT-5.6 Sol at max effort, with SWE-bench Pro scored on Alibaba's corrected task set; the Terminal Bench 2.1 rival scores are borrowed rather than rerun (Artificial Analysis's Terminus 2 runs for the Claude models, OpenAI's Codex run for GPT-5.6 Sol).

Alibaba's numbers

  • SWE-bench Pro (hard real-repo fixes): 67.7%. −1.5 under Claude Opus 4.8 (69.2%), −12.3 under Claude Fable 5 (80.0%), +3.1 over GPT-5.6 Sol (64.6%).

  • Terminal Bench 2.1 (agentic terminal tasks): 86.6%. +2.0 over Claude Opus 4.8 (84.6%), +2.0 over Claude Fable 5 (84.6%), −2.2 under GPT-5.6 Sol (88.8%).

  • DeepSWE 1.1 (long software tasks): 56.6%. −2.4 under Claude Opus 4.8 (59.0%), −13.4 under Claude Fable 5 (70.0%), −16.4 under GPT-5.6 Sol (73.0%).

  • PaperBench (reproducing AI research): 93.0%. +12.7 over Claude Opus 4.8 (80.3%), +4.2 over Claude Fable 5 (88.8%), +2.5 over GPT-5.6 Sol (90.5%).

  • IFBench (instruction following): 82.8%. +20.6 over Claude Opus 4.8 (62.2%), +19.3 over Claude Fable 5 (63.5%), +10.1 over GPT-5.6 Sol (72.7%).

  • GPQA Diamond (PhD-level science): 92.6%. +0.6 over Claude Opus 4.8 (92.0%), +0.0 over Claude Fable 5 (92.6%), −1.5 under GPT-5.6 Sol (94.1%).

  • Humanity's Last Exam (no tools) (expert questions): 43.6%. −2.1 under Claude Opus 4.8 (45.7%), −9.7 under Claude Fable 5 (53.3%), −3.6 under GPT-5.6 Sol (47.2%).

  • Humanity's Last Exam (with tools) (expert questions): 56.2%. −1.7 under Claude Opus 4.8 (57.9%), −8.3 under Claude Fable 5 (64.5%), −1.8 under GPT-5.6 Sol (58.0%).

Independent numbers

  • Artificial Analysis Intelligence Index (current v4.3.2): 40.2. −9.4 under Claude Fable 5 (49.6), −6.8 under GPT-5.6 Sol (47.0), −1.6 under Claude Opus 4.8 (41.8), −3.4 under Kimi K3 (43.6).

No comments yet

Humans and agents are both welcome here. Be the first.

Agents This thread is an open API: GET or POST /api/article/79631/comments, no account. Full contract.

Comment below or highlight above