Alibaba releases 2.4T-parameter Qwen3.8-Max, open-sources next week

On August 3, Alibaba's Qwen team released the flagship model Qwen3.8-Max. Featuring a total of 2.4T parameters, this MoE architecture supports a 1 million token context and can process text, image, and video inputs. The company emphasized that it sets a new benchmark in autonomous coding and collaborative work, claiming it can independently evolve and develop from an empty folder for over 10 days without human intervention. The model scored 1668 points in the Frontend Code Arena, ranking fourth and closely trailing Claude Opus 5 (Max) and Kimi K3. API pricing is set at $2 per million input tokens and $6 per million output tokens. Officials announced that the weights for both Qwen3.8-Max and Qwen3.8-27B will be open-sourced next week, with inference providers like vLLM, TokenSpeed, and Baseten confirming Day-0 support. This marks the industry's second open-weight model exceeding 2T parameters, following Kimi-K3.

Confirmed

• Model release: Qwen3.8-Max officially launched, featuring 2.4T parameters, MoE architecture, 1 million token context, with a focus on coding and collaboration.
• Open-source plan: Weights for Qwen3.8-Max and Qwen3.8-27B to be open-sourced next week.
• Benchmark performance: Scored 1668 in the Frontend Code Arena, ranking 4th.
• Pricing: $2/million input tokens, $6/million output tokens.
• Ecosystem support: vLLM, TokenSpeed, and Baseten announce Day-0 support.

Unconfirmed

• Specific details regarding its autonomous coding capabilities (such as evolving for over 10 days) originate from forwarded posts and have not been detailed by the official team.
• The open-source license type has not yet been announced, though some users are hoping for an MIT license.

Why it matters

This is Alibaba's first open-source Max-tier model and the second globally to exceed 2T parameters after Kimi-K3, signaling that Chinese open-source models are approaching US SOTA levels in both scale and capability. Its autonomous programming abilities and low cost (rumored to slash costs by 80%) could potentially reshape the coding tools market.

2026-08-03 ~ 2026-08-03 · 30 related posts

• Alibaba Releases Qwen3.8-Max: A New Bar for Coding and Cowork — Alibaba_Qwen · 2026-08-03
• Alibaba's Qwen3.8-Max Released, TokenSpeed Announces Inference Optimization — zhyncs42 · 2026-08-03
• [source] Alibaba Launches 2.4T Parameter Qwen3.8-Max, Open Weights Coming Next Week — Alibaba_Qwen · 2026-08-03
• [source] Qwen3.8-Max Hits Text Arena: 2.4T Parameters Focused on Autonomous Coding — Alibaba_Qwen · 2026-08-03
• Qwen Open-Sources First Max-Class Model, Exceeding 2T Parameters — cedric_chee · 2026-08-03
• Users Call for 200B Parameter Open-Source Release from Qwen — adrianscottcom · 2026-08-03
• Alibaba launches Qwen3.8-Max at 2.4T parameters, excelling in autonomous coding — cephaloform · 2026-08-03
• Qwen3.8-Max Hits #4 on Frontend Code Arena, Reshaping Cost-Performance Frontier — Scobleizer · 2026-08-03
• Alibaba to Open-Source Qwen3.8-Max with 2.4T Parameters and Autonomous Coding — EugeneWare · 2026-08-03
• Alibaba Announces Qwen3.8-Max: 2.4T Parameters and 1M Context Window — Scobleizer · 2026-08-03
• Alibaba's Qwen Announces Open Weights for Qwen3.8-Max and 27B Next Week — multimodalart · 2026-08-03
• Alibaba Launches 2.4T Parameter Qwen3.8-Max: Rivals Claude in Performance at a Fraction of the Cost — 量子位 · 2026-08-03
• vLLM Announces Day-0 Support for Open-Source Qwen3.8-Max — AccBalanced · 2026-08-03
• Alibaba Releases Qwen3.8-Max: 2.4T Params, Open-Sourcing Next Week — APPSO · 2026-08-03
• Alibaba Cloud Officially Releases Qwen3.8-Max Model — op7418 · 2026-08-03
• Report: Alibaba's Qwen3.8-Max to Open Weights, Autonomously Codes for 16 Days at 80% Lower Cost — AccBalanced · 2026-08-03
• Qwen3.8-Max Released: 2.4T Parameters, Competitive Pricing, and Surprising Writing Skills — 数字生命卡兹克 · 2026-08-03
• Baseten to Serve Alibaba's Qwen3-Max Model — baseten · 2026-08-03
• Alibaba Launches 2.4T-Parameter Qwen3.8-Max, Open Weights Next Week — Deep_Ladder_4679 · 2026-08-03
• Qwen 3.8 Max: 2.4T-parameter multimodal model, open-weight expected — Prompt Engineering · 2026-08-03

10 near-duplicate retellings: yacineMTB · airesearch12 · 千问大模型 · xiaohu · AccBalanced · 智东西 · victormustar · haider1 · Alibaba_Qwen · baseten