Morning Edition · Tuesday, August 4, 2026Published at 2:03 AM EDT · New York
Alibaba Ships Qwen3.8-Max, a 2.4-Trillion-Parameter Model It Promises to Open-Weight Next Week
The model activates 95 billion parameters per request and prices at two dollars per million input tokens. It still trails Anthropic's Fable 5 on the hardest public software-engineering benchmark.
Alibaba released Qwen3.8-Max on Monday, a mixture-of-experts model with 2.4 trillion total parameters, roughly 95 billion of which activate on any single request. It handles text, images and video, and it supports a context window of up to one million tokens. Alibaba serves it through QwenCloud at two dollars per million input tokens and six dollars per million output tokens, with cached input at twenty-five cents. The company said the weights will be published on Hugging Face and ModelScope next week under an Apache 2.0 license, alongside a much smaller Qwen3.8-27B, which MarkTechPost reports would make it the first downloadable Max-class Qwen model.
The company's own numbers put the model ahead of OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 on several coding, multimodal and engineering evaluations, and behind them on parts of general reasoning. The actual difference is narrower than the marketing suggests. Qwen3.8-Max reports 67.7 on SWE-bench Pro, while Anthropic claims 80.3 on the same benchmark for Fable 5 and Mythos 5. Every one of these figures comes from a vendor rather than from an independent reproduction, and the weights that would allow outside verification have not been released.
The independent signal so far is informal. A team at Atomic Chat gave Qwen3.8-Max and Fable 5 the same task: three self-contained interactive HTML scenes with three-dimensional physics, including marbles riding a wheel through loops. The team reported that the Qwen output was better and cost close to seven times less. That is one prompt run by one team, which is evidence about price and performance rather than proof of capability parity.
The presentation around the release is as deliberate as the model itself. Alibaba's promotional video shows the model spending hours on chip design work before moving to biology tasks, a pairing that Russian-language AI coverage read as a direct answer to United States export controls on semiconductors. Investors responded to the pricing rather than to the benchmark results. Alibaba's Hong Kong-listed shares closed about 7% higher at HK$125.20, and its American depositary receipts rose in premarket trading.
- If true, who benefits
Alibaba's cloud division, which uses a cheap flagship model to acquire customers, and Chinese policymakers who want evidence that United States export controls have not stopped frontier work, while OpenAI and Anthropic lose per-token pricing power and holders of the Hong Kong line gain from a close of HK$125.20, up 7.01%.
- The nuance
The release, price and share move are independently confirmed, but the weights had not appeared when the stock rose, and every capability figure is Alibaba's own: independent trackers place the 67.7 on SWE-bench Pro mid-pack, ahead of GPT-5.6 Sol and behind both Opus 4.8 and Fable 5, so parity rests on the subset of tests Alibaba chose to publish, and the return to open weights follows a year of keeping flagship Qwen models closed.
An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.
What this means
A downloadable Max-class model at two dollars per million input tokens competes with the closed labs at their weakest point, which is the price of the second-best answer. Enterprises that route bulk agentic and coding traffic through an application programming interface (API) now have a self-hostable option at a fraction of frontier pricing. That pressures per-token margins at OpenAI and Anthropic and strengthens Alibaba's cloud business, where the model works as a customer-acquisition mechanism rather than as the product. The unresolved question is whether the released weights reproduce the claimed scores. If independent runs come in near 67.7 on SWE-bench Pro, Alibaba has a strong value model that still trails the frontier on hard software work. If they come in well below, the vendor numbers were tuned and price is the only real advantage.
What to watch
- Whether the promised Hugging Face and ModelScope weights actually appear next week, and whether outside groups reproduce the coding scores. A gap between the vendor table and independent runs would be the clearest evidence yet that Max-class benchmark claims need discounting.
- How OpenAI and Anthropic respond on price rather than on capability. A repricing of their mid-tier coding endpoints would confirm that open Chinese weights now set the lowest price enterprises expect to pay per token.
- Whether Western enterprises and governments allow a Chinese-origin open-weight model into production stacks, since self-hosting removes the data-residency objection but not the procurement politics.
Observations to monitor, not financial advice.
Synthesized from: Polylog editors · CNBC · MarkTechPost
Part of a tracked trend
Chinese Labs Reach Frontier Parity
Chinese labs increasingly match or beat United States frontier offerings on independent benchmarks across modalities, competing on closed metered APIs as well as open weights.
More from this edition
- OpenAI Publishes Lean-Verified Proofs From Its Astra Model for Ten Long-Open Mathematics Problems
- Tencent's Hyra Agent Reports Beating the Historical Best on 29 of 55 Open Mathematics Problems
- An Open-Source Runtime Puts an 80-Billion-Parameter Qwen Model on a Mac in 4.3 Gigabytes of Memory
- Meta Puts Muse Image and a Muse Video Preview Into Its Consumer Apps With Conversational Editing
- Meta's Open Vision Models Move Into a $41.5 Million Robotic Wheelchair Program and Berkeley Lab Science
- AI Evaluation Turns Into a Product Market as Researchers Show Cheap Open Models Can Grade Proofs
- New Benchmarks Target the Decision Agents Make Before They Answer
- Agent Observability Reaches Cloud Consoles as Enterprises Report Measured Deployment Results
- Researchers Propose a Zero-Trust Registry for Agent Skills After Finding Claims Do Not Match Capabilities
- A Self-Refining Agent Takes On OpenFOAM Configuration, One of Engineering's Reliable Time Sinks
- A Paper Revives Leibniz, Turing and Searle to Argue Consciousness Tests Belong in AI Safety Work
Comments
0No comments yet.