Qwen3.8 2.4T A95B

Qwen3.8 2.4T A95B is positioned as an open-weight sparse mixture-of-experts model in the Qwen family, activating 95 billion of its 2.4 trillion total parameters per token. Alibaba released the open weights, framing it as the open-weights counterpart to Qwen3.8-Max and aiming to bring near-frontier capability into the open ecosystem. The architecture is designed for heavy lifting tasks: coding, research, complex reasoning, and agentic workflows where a single response can chain tool use, multi-step planning, and long-context synthesis.

In practical deployment, the model has already been documented for serving on NVIDIA GB300 NVL72 hardware with a configurable reasoning mode, signaling that operators can dial reasoning effort up or down depending on the workload. DeepInfra lists a hosted endpoint for text generation under the Qwen namespace, giving developers an inference option without standing up their own GPU fleet. The combination of open weights, very large total capacity with moderate per-token activation, and configurable reasoning points to a model aimed at teams that want frontier-style behavior but retain the ability to self-host, fine-tune, and control cost on large-context jobs.

Charm Hyperqwen3.8-2.4t-a95bqwen

Open provider page Provider docs

Quick Info

Provider: Charm Hyper
Model key: qwen3.8-2.4t-a95b
Release date: Aug 26, 2026
Last updated: Aug 26, 2026
Input modalities
Output modalities
Capabilities

Cost

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

Limits

Output tokens: 128,000 tokens
Context window: 1,000,000 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

Charm Hyper

CoverageBenchmark

A Sep 7, 2026 recap article documents the Qwen3.8-Max rollout timeline: Alibaba's API debut on August 2, 2026, followed by the open-weight Qwen3.8-2.4T-A95B release on Hugging Face and ModelScope on August 13, 2026, and the dense Qwen3.8-27B sibling on August 14, 2026. It reiterates the 2.4 trillion-parameter MoE archi The article positions Qwen3.8-Max as targeted at coding agents and multi-step software engineering tasks rather than general chat, noting the model's combination of scale and sparse activation makes self-hosting feasible for teams with serious infrastructure. It does not introduce original benchmark data, instead refer

Charm Hyper

Coverage

NVIDIA's technical blog (Aug 12, 2026) details day-0 serving of Qwen3.8-2.4T-A95B on the GB300 NVL72 platform, confirming the model's 2.4 trillion total / 95 billion active MoE architecture with hybrid full and linear attention, up to 1M-token context, and up to 128K-token output. Without additional tuning, the blog re The post positions Qwen3.8-2.4T-A95B as Alibaba's largest open-weight release and frames NVIDIA's role as a serving-stack collaborator rather than model creator, emphasizing multinode deployments through optimized kernels and distributed serving recipes for demanding reasoning and agentic workloads. While the blog surf

Charm Hyper

CoverageDiscourse

A Hugging Face discussion thread under the Qwen/Qwen3.8-2.4T-A95B repository surfaces the official model card text, which states that Qwen3.8-Max is the official service version based on Qwen3.8-2.4T-A95B with additional features: vision input, non-thinking support, 1M context length by default, and official built-in t This thread provides useful first-party documentation context for distinguishing the Qwen3.8-2.4T-A95B open-weight artifact from the Qwen3.8-Max API/service variant, which matters for developers deciding what features they actually get by self-hosting versus using the managed API. The discussion also flags industry con

Charm Hyper

CoverageBenchmark

36Kr's report, sourced from the Alibaba Qwen team's announcements, confirms the open-source release of Qwen3.8-2.4T-A95B—the first time Alibaba has publicly opened the weights of a Max-class flagship model—on both Hugging Face and ModelScope. The article names the exact variant and specifies a total of 2.4 trillion par The same report adds deployment and operational details: Qwen3.8-2.4T-A95B can be served through SGLang, vLLM, and TokenSpeed inference engines using the latest Qwen3.8 recipes with parallel strategies tuned to weight precision and GPU configuration, and it runs in thinking mode by default with no option to disable it.