Qwen3.8-2.4T-A95B Model Card (Qwen Team, August 2026)

The Qwen3.8-2.4T-A95B model card is the Hugging Face repository page under which Alibaba's Qwen team released the weights of Qwen3.8-Max, the first Qwen-Max-class model to be opened (Source: huggingface.co). The repository was created on 8 August 2026 and last modified on 12 August 2026; the card itself is undated and cites the team's launch post of 2 August 2026, Qwen3.8-Max: A New Bar for Coding and Cowork, as its reference and for citation. The card distinguishes the open checkpoint from the hosted Qwen3.8-Max service, which it describes as "the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools." Its benchmark table reproduces the launch post's "Full Benchmark Table" and labels the model column "Qwen3.8-Max".

Model description

The card describes Qwen3.8 as built on the architectural foundation of Qwen3.5, with gains in coding, professional work, research and long-horizon agentic tasks, stronger autonomous planning and handling of environment feedback, broader harness and tool compatibility, and a reasoning_effort control with preserve_thinking for retaining reasoning context across turns. The stated architecture is a causal language model of 2.4 trillion total and 95 billion activated parameters: 92 layers arranged as 23 blocks of three Gated DeltaNet → MoE layers followed by one Gated Attention → MoE layer; Gated DeltaNet with 128 linear-attention heads for V and 16 for QK at head dimension 128; gated attention with 64 query heads and 4 key-value heads at head dimension 256 with a 64-dimensional rotary embedding; a mixture of 512 experts activating 10 routed plus 1 shared, each with intermediate dimension 2048; a padded vocabulary of 248,320; and multi-token prediction trained with multiple steps. Native context is 262,144 tokens, extensible to 1,010,000. The weights are distributed in BF16 as a 2.4T-parameter safetensors checkpoint for vLLM, SGLang and TokenSpeed.

Evaluation approach and findings

The "Benchmark Results" table compares Qwen3.8-Max with Claude Opus 4.8, Claude Fable 5, GPT-5.6 Sol at maximum effort, and the team's previous Qwen3.7-Max across three groups: coding agents (Terminal Bench 2.1, SWE-bench Pro, DeepSWE 1.1, NL2Repo-Bench, FrontierSWE, MLS-Bench-Lite, PaperBench, AndroidBench, and four in-house benchmarks — QwenSWEBench, QwenQoderBench, QwenReactBench and QwenSVGBench); general agents (CoWorkBench, WorkSpaceBench, JobBench, SkillsBench, Agents' Last Exam, Automation-Bench, Toolathlon Verified, WideSearch and HLE with tools); and general capabilities (GPQA Diamond, HLE, IFBench, $OneMillion-Bench, HealthBench, PLawBench, PRBench-Legal and -Finance, MRCR v2 at 256K and LongBench v2).

Twenty footnotes condition the numbers. Most agentic results are the Qwen team's own runs through the Claude Code harness with a 256K context, often with stated timeouts, sampling parameters and run counts (for example avg@10 with a 5-hour timeout on Terminal Bench 2.1; avg@3 over up to 12 hours per run on PaperBench, judged by Claude Opus 4.6). For Terminal Bench 2.1 the comparison figures are instead "the best published score across harnesses" — Artificial Analysis's Terminus 2 results for the Anthropic models and OpenAI's Codex result for GPT-5.6 Sol — and for FrontierSWE and MLS-Bench-Lite the baselines come from the official leaderboards, so several rows mix harnesses. SkillsBench places the Anthropic models on Claude Code, GPT-5.6 Sol on Codex and the Qwen models on OpenCode, "all results from our own testing." The SWE-bench Pro figures are on a version of the benchmark in which the team says problematic tasks were corrected and all baselines re-run. A first footnote states that "Fable5 results may involve fallbacks," and empty cells denote scores "not yet available or not applicable."

Read against that framing, the card places Qwen3.8-Max above its listed comparators on PaperBench (93.0), IFBench (82.8), HealthBench (60.2), PLawBench (73.2), PRBench-Finance (58.3) and WideSearch (81.9), level with Fable 5 on GPQA Diamond (92.6) and PRBench-Legal (57.6), and below one or more of them on Terminal Bench 2.1 (86.6 against 88.8 for GPT-5.6 Sol), SWE-bench Pro (67.7 against 80.0 for Fable 5), DeepSWE 1.1, HLE without tools (43.6 against 53.3) and the long-context rows. The card reports no safety, refusal or dangerous-capability evaluations, no license text in the body, and no training-data or compute figures.

Provenance and confidence

The raw file is a scrape of the Hugging Face page and includes the site's usage snippets and download widgets around the card text. All figures are the developer's; the table's mixed provenance (own runs, quoted leaderboards and third-party harness results in the same row) means the comparison columns are not uniformly comparable, and no independent replication of the Qwen3.8-Max figures has been identified. Confidence is medium: a single developer-authored source, fast-decay benchmark claims, and a hosted "Qwen3.8-Max" service that differs in features from the open checkpoint the card documents.

Reported results

Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.

Relationships

• supports: Qwen3.8-Max (primary reference for the open-weight release's architecture and benchmark table)
• related: Qwen3 (predecessor generation; the card describes Qwen3.8 as built on Qwen3.5's architecture), Claude Opus 4.8, Claude Fable 5, GPT-5.6 (Sol, Terra, Luna) (comparison models), Terminal-Bench, SWE-Bench Pro, PaperBench, Humanity's Last Exam, GPQA, HealthBench, MRCR
• instance-of: Open-Weight Frontier Models