Qwen3.8 2.4T A95B

Qwen3.8 2.4T A95B is an open-weight sparse mixture-of-experts model created by Qwen, distinguished by a 95 billion active parameter count drawn from a much larger 2.4 trillion total parameter pool. This very sparse activation ratio allows the model to keep inference compute closer to its active footprint while retaining the capacity of a frontier-scale MoE, positioning it as the open-weight counterpart to Qwen3.8 Max. The architecture is purpose-built for workloads that demand substantial reasoning depth, including long-form coding, research synthesis, and multi-step agentic pipelines where tool use and structured outputs are essential.

Released on August 12, 2026, and hosted on Hugging Face under the Qwen organization, the model drew significant community attention shortly after launch. Its intended fit centers on complex reasoning and coding, where the combination of high active capacity and large total parameter headroom supports nuanced problem solving while remaining an open-weight option that teams can self-host. Practically, it is best suited to advanced research assistants and autonomous agent systems that benefit from a sparse MoE design aimed at balancing scale with controllable active compute per token.

AIHubMixqwen3.8-2.4t-a95bqwen

Open provider page Provider docs

Quick Info

Provider: AIHubMix
Model key: qwen3.8-2.4t-a95b
Release date: Aug 12, 2026
Last updated: Aug 12, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $2.00
Output token cost: $6.00

Limits

Output tokens: 262,000 tokens
Context window: 262,000 tokens

Transparent token rates

Compare Qwen3.8 2.4T A95B pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Qwen3.8 2.4T A95B

AIHubMix

CoverageBenchmark

Alibaba shipped Qwen3.8-Max as a sparse mixture-of-experts model with 2.4 trillion total parameters and 95 billion active parameters, first going live through the QwenCloud API on August 2, 2026 and then published as open weights under the name Qwen3.8-2.4T-A95B on Hugging Face and ModelScope on August 13, 2026, with a The article also situates Qwen3.8-2.4T-A95B within a broader September 2026 landscape of competing frontier models, arguing that Alibaba's choice to hand over full weights is a notable counterpoint to labs that have kept capabilities behind closed APIs. It cites a CheapestInference breakdown for the August 13/14 rollou

AIHubMix

Coverage

NVIDIA's developer blog dated August 12, 2026 documents the open-weight release of Qwen3.8-2.4T-A95B, describing it as Alibaba's largest open-weight model with 2.4T total parameters, 95B activated per token, a fine-grained MoE architecture, a hybrid full and linear attention design, a context window of up to one millio The post reports Day-0 throughput of over 4,000 tokens per second per GPU and over 350 tokens per second per user in FP8 on NVIDIA GB300 NVL72, with further gains expected from NVFP4 optimizations. It also details that NVIDIA NeMo AutoModel supports post-training via full SFT or memory-efficient LoRA directly on Huggin

AIHubMix

CoverageDiscourse

A Hugging Face community discussion opened by NodeLinker on August 12, 2026 directly addresses the Qwen/Qwen3.8-2.4T-A95B repository and documents that the released open weights are text-only and have been stripped of capabilities present in the cloud-hosted Qwen3.8-Max, specifically vision input and the 1M-token defau The discussion expresses developer concern that stripping vision and the full 1M context from the open weights reduces the model's utility for self-hosters and contrasts this with Kimi K3's full open release. It also raises the prospect of Qwen3.8-27B open weights arriving the following week with similarly gated featur

AIHubMix

CoverageBenchmark

BenchLM's tracker page, current as of September 10, 2026, explicitly maps the "Qwen3.8 Max" profile to the Qwen3.8-2.4T-A95B model card and records the release date as August 3, 2026. It catalogs the model as an open-weight, reasoning-capable design with a 1M-token context window and reports an overall capability score The page also breaks down category percentiles for Qwen3.8-2.4T-A95B: Agentic at the 95th percentile (rank 9/152), Coding 87th (20/151), Knowledge 91st (17/183), Multimodal 91st, and Instruction Following 86th, while noting Reasoning, Math, and Multilingual are not rank-eligible. Speed is listed at 38 tok/s with a 55.3