Qwen ships Qwen3.8-2.4T-A95B, the first open Qwen-Max-class model
Qwen has published Qwen3.8-2.4T-A95B on Hugging Face, the open-weight checkpoint behind the new Qwen3.8 model generation. The card states Qwen3.8 is the first Qwen-Max-class model the team has released openly, following what it describes as widespread community adoption of the Qwen3.5 and Qwen3.6 series. Built on the architectural foundation of Qwen3.5, Qwen3.8 targets substantial gains in coding, professional work, research, and long-horizon agentic tasks, with an explicit design goal of carrying complex, multi-step tasks through to completion more reliably, not just answering harder questions.
The released checkpoint has 2.4T parameters in total with 95B activated per token, 92 layers, a hidden dimension of 8192, and a token embedding size of 248,320 (padded). Its Mixture-of-Experts layer has 512 experts, of which 10 routed plus 1 shared are active per token, with an expert intermediate dimension of 2048. The architecture combines Gated DeltaNet blocks (128 attention heads for V, 16 for QK) with Gated Attention blocks (64 heads for Q, 4 for KV, a 256 head dimension, and 64-dimensional rotary position embeddings) in a repeating layout of 23 groups, each made of three DeltaNet-to-MoE blocks followed by one Gated-Attention-to-MoE block; the model also uses multi-token prediction trained over multiple steps. Native context length is 262,144 tokens, extensible up to 1,010,000 tokens.
Qwen3.8-2.4T-A95B is text-only: multimodal input is not supported, and thinking mode cannot be turned off, so every response begins with a reasoning block before the final answer. Reasoning depth is tunable through a reasoning_effort parameter with three levels: xhigh (the default, for thorough analysis), medium (balancing accuracy and speed), and low (optimized for speed and cost); preserve_thinking, which retains reasoning context from earlier turns in a conversation, is enabled by default. The card recommends temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0 and repetition_penalty=1.0 for generation. The weights are compatible with Transformers, vLLM, SGLang and TokenSpeed, and can be served directly with Docker Model Runner or through HuggingChat, Colab and Kaggle notebooks.
For teams that do not want to self-host, Qwen offers a hosted counterpart, Qwen3.8-Max, through the Qwen Cloud API. Qwen3.8-Max is described as the same base model with more features layered on: vision input, support for turning thinking off, a 1M-token context length by default, and official built-in tools. No pricing for Qwen3.8-Max or the API service is given in the material.
The card reports evaluation across a long list of coding- and agent-focused benchmarks, among them Terminal Bench 2.1, SWE-bench Pro, DeepSWE 1.1, NL2Repo-Bench, FrontierSWE, MLS-Bench-Lite, PaperBench, AndroidBench, and several in-house suites (QwenSWEBench, QwenQoderBench, QwenReactBench, QwenSVGBench, CoWorkBench, SkillsBench, Automation-Bench and WideSearch), most of them run through the Claude Code harness. The named comparison points are Claude Opus 4.8, Claude Fable 5 (with the Terminus 2 harness) and GPT-5.6 Sol (with Codex). The retrieved material lists the evaluation methodology and footnotes for each benchmark but does not include the resulting score table, so no numeric outcome of these comparisons can be reported here.
Key facts
- Qwen3.8-2.4T-A95B has 2.4T total parameters with 95B activated per token, built on the Qwen3.5 architecture.
- It is the first Qwen-Max-class model released with open weights, aimed at coding, professional work, research and long-horizon agentic tasks.
- Native context length is 262,144 tokens, extensible to 1,010,000 tokens; the model is text-only and thinking mode cannot be disabled.
- Reasoning depth is tunable via reasoning_effort (xhigh, medium or low), and preserve_thinking keeps reasoning context across turns by default.
- A hosted variant, Qwen3.8-Max, adds vision input, non-thinking mode, a 1M-token default context length and built-in tools through the Qwen Cloud API.
Why it matters
Qwen-Max has previously been the closed, API-only tier of the Qwen lineup. Qwen3.8-2.4T-A95B is, by the card's own description, the first Qwen-Max-class model released with downloadable weights rather than kept behind an API. It carries the jump to 2.4T total parameters with 95B active, sits on the Qwen3.5 architectural base, and is pitched specifically at coding, professional and research work and at long-horizon agentic tasks, tasks that require finishing a multi-step job reliably rather than just producing a single correct answer.
Who it affects
The audience is developers and teams who want frontier-scale open weights to self-host, fine-tune or inspect directly, rather than casual chat users. Running a 2.4T-parameter model (95B active) requires serious infrastructure, and the card repeatedly frames evaluation and usage around agentic coding harnesses such as Claude Code, SGLang and vLLM, which points at builders of coding agents and automation tooling as the primary users. Those who want a lighter-weight, fully managed option are pointed to the hosted Qwen3.8-Max instead.
How to use it
The weights can be loaded directly through Hugging Face Transformers, or served with vLLM, SGLang, TokenSpeed or Docker Model Runner; the card gives copy-paste code for each. Reasoning depth is controlled with reasoning_effort, set to xhigh by default and adjustable to medium or low, and preserve_thinking (also default-on) keeps prior reasoning in context across turns. Recommended sampling settings are temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0 and repetition_penalty=1.0. Thinking cannot be switched off on this checkpoint; for non-thinking mode, vision input, a 1M-token default context and built-in tools, the card points to the hosted Qwen3.8-Max via the Qwen Cloud API, for which no pricing is given.
How solid is it
The claims come from Qwen's own model card, a first-party document, not an independent review. It describes evaluation across a long roster of coding and agentic benchmarks, including Terminal Bench 2.1, SWE-bench Pro, DeepSWE 1.1, NL2Repo-Bench, FrontierSWE and several in-house suites, mostly run through the Claude Code harness, with Claude Opus 4.8, Claude Fable 5 and GPT-5.6 Sol named as comparison models. The material available here is the methodology and footnotes for those benchmarks; the actual score table was not present in the retrieved text, so this retelling cannot report how Qwen3.8-2.4T-A95B actually performed against those comparisons.
Risks and caveats
Thinking mode is mandatory and cannot be disabled on this checkpoint, so every response carries a reasoning preamble, adding latency and token cost even for simple queries. The open-weight release is text-only; multimodal input is reserved for the hosted Qwen3.8-Max. The card gives no release date, no individual author or team credit (it speaks only as "we" and "Qwen"), and no pricing for the hosted API. Self-hosting a 2.4T-parameter model, even with 95B active, is a heavy hardware commitment that most teams will not be able to take on directly.
“Qwen3.8-2.4T-A95B is a text-only model that requires thinking mode for all interactions.”
— the model card (Qwen)