ainewsblitz.com

Breaking

Alibaba's Qwen3.8-Max Takes BenchLM's Open-Weight Lead at 78.9

  • Foundation Models
  • AI Agents
  • Open Source

Alibaba's Qwen3.8-Max has taken the top spot in BenchLM's open-weight ranking with a score of 78.9, while Qwen has published Max-class weights under the name Qwen3.8-2.4T-A95B. The release puts a 2.4-trillion-parameter mixture-of-experts model, with 95 billion parameters active per token, into the hands of developers evaluating frontier-scale systems. Alibaba announced the hosted Qwen3.8-Max on Aug. 3, 2026, describing it as its largest and most capable Qwen model to date, with capabilities spanning coding, professional work, research, long-horizon tasks and visual intelligence. The model was initially offered through Alibaba Cloud Model Studio, with weights scheduled to follow in the company's announcement. The Qwen3.8 repository calls the release the first Qwen-Max-class model made openly available and says it builds on the Qwen3.5 architecture. On Hugging Face, the package is a post-trained Transformers checkpoint compatible with vLLM, SGLang and TokenSpeed. Its model card distinguishes the text-only Qwen3.8-2.4T-A95B checkpoint from the feature-complete Qwen3.8-Max service: the hosted version adds vision input, non-thinking support, built-in tools and a 1M-token context by default, while the open checkpoint supports text only, requires thinking mode and has 262,144 tokens of native context that can be extended to 1,010,000. BenchLM's open-weight ranking lists Qwen3.8 Max first among 102 models at 78.9, ahead of Qwen3.8-27B at 72.4, dots3-note Preview at 68.8, Ornith-1.5-397B at 68.5 and MiniMax M3 at 68.4. That result is narrower than a first-place finish across the entire field: BenchLM's model profile places Qwen3.8 Max sixth out of 225 models on its broader public leaderboard, with a score of 79.04, and shows 54 source-displayable rows across 404 tracked benchmark slots. Qwen's published comparison table reports 86.6 on Terminal Bench 2.1 for Qwen3.8-Max, versus 84.6 for Opus 4.8 and 74.5 for Qwen3.7-Max. It also reports 93.0 on PaperBench against 80.3 for Opus 4.8 and 81.9 on WideSearch against 72.9. The picture is not uniformly dominant: Qwen3.8-Max scores 67.7 on SWE-bench Pro against 69.2 for Opus 4.8, and 56.6 on DeepSWE 1.1 against 59.0. The model card notes that Qwen's Terminal Bench result used Claude Code with an average over 10 runs, while comparator figures were drawn from the best published results across different harnesses, so the numbers are informative rather than one fully controlled head-to-head test. Technically, the checkpoint contains 512 experts, with 10 routed experts plus one shared expert activated for each token, across 92 layers. The official Qwen3.8-Max README lists API pricing of $2 per 1 million input tokens, $6 per 1 million output tokens and $0.25 for implicit caching; it also gives maximum input and output limits of 991.80K and 131.07K tokens. Self-hosting remains a major engineering undertaking: a vLLM deployment guide estimates approximately 4.89 TB of storage and weight memory for the FP16/BF16 checkpoint. Early developer feedback is positive on coding and long-running work but uneven on integration. In a discussion of the preview, one user reported completing as many private-repository tasks as Fable, while throughput ranged from 20 to 80 tokens per second and skill loading was a noticeable difference; other users described OpenCode errors and timeouts in the same developer discussion. A separate Qwen GitHub issue records a preview session that stopped mid-generation after roughly two to three minutes on a long Python request and remained stuck. These reports concern the preview and early integrations rather than controlled evaluations, but they underline latency, harness compatibility and recoverability as practical constraints. Qwen3.8-Max therefore arrives with a mixed evidence profile: a leading open-weight aggregate and standout results on selected reasoning, research and agent tests, alongside weaker scores on other coding evaluations and significant serving requirements. For adopters, the central question is whether the model's benchmark strengths, latency and cost can be reproduced on their own workloads.