Source: Alibaba Group — 2026-08-03
Summary
Alibaba announced Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model that activates only about 95 billion parameters per token, with a 1-million-token context window and multimodal (text, image, video) input. Alibaba reports 86.6 on Terminal-Bench 2.1 and a leading 93.0 on PaperBench, though as of this writing those numbers are self-reported and haven't yet been independently reproduced by third-party benchmark trackers like Artificial Analysis or LMArena. Open weights are slated for Hugging Face the week of August 10, 2026.
Key Takeaways
- 2.4T total parameters but only ~95B activated per token — a sparsity ratio of roughly 25:1, which is what keeps inference cost and latency manageable at this scale.
- 1-million-token context window plus multimodal input puts it in the same tier as the biggest frontier context windows currently shipping.
- Self-reported scores: 86.6 on Terminal-Bench 2.1, 93.0 on PaperBench (leading that benchmark) — worth treating as a claim, not a verified result, until independent evals land.
- Open weights on Hugging Face means anyone can run their own eval rather than trusting the vendor's numbers — a real test of the claim is imminent.
Reel Script
Hook (17s)
Alibaba just claimed the top score on a research-paper benchmark with a model 2.4 trillion parameters in size — but here's the catch: almost none of those parameters are doing work at any given moment.
Core Concept (95s)
Qwen3.8-Max uses something called mixture-of-experts, or MoE. Picture a giant hospital with 2.4 trillion "doctors" on staff, but for any single patient, only a small specialist team of about 95 billion of them actually gets called into the room — the rest stay idle. That's the trick: the model has a massive total capacity, but a routing mechanism decides which small slice of the network actually processes each token, so you get the knowledge of a huge model with the compute cost closer to a much smaller one. That ratio here — 95 billion active out of 2.4 trillion total — is roughly 25-to-1, which is aggressive even by current MoE standards. On top of that, it ships with a 1-million-token context window, meaning it can hold roughly a small book's worth of text in working memory at once, and it takes image and video input too, not just text.
Hands-On (90s)
The benchmark numbers Alibaba is publishing: 86.6 on Terminal-Bench 2.1, which tests an agent's ability to actually complete tasks in a real terminal environment, and 93.0 on PaperBench, where it's claimed as the current leader — PaperBench tests whether a model can reproduce the results of a research paper from scratch. Those are strong claims. But here's the honest caveat worth saying out loud on camera: these are Alibaba's own reported numbers. As of this release, neither Artificial Analysis nor LMArena — the independent trackers most people actually trust — have reproduced them yet. That's not a knock on the model, it's just how self-reported benchmarks always start. The real test comes the week of August 10, when Alibaba open-sources the weights on Hugging Face — at that point, anyone can run the eval themselves and either confirm or debunk the number.
Takeaway (28s)
Big MoE claims are cheap until the weights are public — the open release next week is the actual verdict, not this announcement. If you're evaluating models for a research-heavy or long-context workload, bookmark this one and wait a week for independent numbers before betting production traffic on Alibaba's self-reported score.