Qwen3.8 2.4T A95B

Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters, first went live through Alibaba's QwenCloud API on August 2, 2026, with the open-weight variant Qwen3.8-2.4T-A95B following on August 13 on Hugging Face and ModelScope, and a smaller dense sibling Qwen3.8-27B shipping open The article situates the Qwen3.8-2.4T-A95B open-weight release within a crowded September for frontier AI, noting OpenAI, Anthropic, and Google all shipped new models within days of each other. The sparse MoE design, with only 95 billion parameters active on any forward pass, is highlighted as what makes self-hosting f

"Qwen3.8-Max," announced by Alibaba on August 3, 2026, is the company's largest model to date, with 2.4 trillion total parameters. The AI-Driven Lab guide on note.com explains the relationship between the three Qwen3.8 variants — Max, 2.4T-A95B, and 27B — detailing the underlying architecture that combines MoE with hyb The guide emphasizes license nuance, characterizing the release as "open weight but not unconditionally free," and is aimed at engineers and product managers evaluating the Qwen3.8 series for implementation. As an auto-translated explainer it carries an explicit accuracy/nuance caveat, and an active-parameter figure of

WhatLLM's model page for Qwen3.8 2.4T A95B confirms the architecture as 2.4T total / 95B active parameters, a native 262,144-token context extensible to 1,010,000 tokens, mandatory thinking mode with low, medium, and xhigh effort levels, and weights published on the official Qwen Hugging Face organization. Independent Observed deployment metrics across hosted endpoints include fastest observed throughput of 125.1 tokens per second and a lowest observed time-to-first-token of 0.648s, with listed pricing around $1.90 per million input and $5.70 per million output tokens. The analysis positions A95B as an open-weight governance and cus

Alibaba released the open weights for Qwen3.8-2.4T-A95B (also called Qwen3.8-Max), its largest open-weight model, featuring 2.4 trillion total parameters with 95 billion activated per token in a fine-grained mixture-of-experts architecture with hybrid full and linear attention, a context window of up to one million tok The post details the deployment ecosystem around Qwen3.8-2.4T-A95B: NVIDIA NeMo AutoModel supports post-training with full supervised fine-tuning or memory-efficient LoRA fine-tuning directly on Hugging Face checkpoints, while open-source inference recipes are available for SGLang, vLLM, and NVIDIA Dynamo. The model ca

A Hugging Face community discussion on the Qwen/Qwen3.8-2.4T-A95B model repository highlights that the released open weights are text-only and lack several features available in the hosted Qwen3.8-Max variant, notably vision input and the native 1-million-token context window. The thread quotes the official model card The discussion raises concerns about feature gating between the open-weight release and the managed Qwen Cloud deployment, noting that vision-language capabilities and extended context are reserved for the closed Max version. While the open Qwen3.8-2.4T-A95B weights remain useful for general tasks and distillation, the

Alibaba's Qwen team open-sourced the Qwen3.8-2.4T-A95B model, marking the first time the company has publicly released weights for a Max-level flagship model. The model has 2.4 trillion total parameters with 95 billion activated per token, natively supports 262,144 tokens of context extendable to 1.01 million tokens, a The model adopts the underlying architecture of Qwen3.5 and achieves performance improvements in programming, office work, scientific research, and long-cycle agent tasks, deployable through SGLang, vLLM, and TokenSpeed inference engines with framework-specific parallel strategies based on weight precision and GPU conf