FULL STORY

Alibaba Open-Sources Qwen3.8-Max; Ecosystem Rushes to Support It

Alibaba open-sourced its 2.4-trillion-parameter flagship Qwen3.8-Max, and within days Unsloth shrank it by 91% via quantization while vLLM, Together AI and other platforms rolled out hosting support.

2026-08-12 ~ 2026-08-15 · 5 episodes · 34 posts

Episode 1 · Alibaba's Qwen Open-Sources Qwen3.8-Max, a 2.4T-Parameter Flagship MoE Model (2026-08-12, 21 posts)

Alibaba's Qwen team has officially released and open-sourced its flagship model Qwen3.8-Max (Qwen3.8-2.4T-A95B), a 2.4-trillion-parameter mixture-of-experts model activating roughly 95B parameters per token. It is Qwen's largest model to date, its first multimodal model beyond one trillion parameters, and the first time the team has opened Max-tier weights. The model topped Hugging Face's trending chart shortly after release and received Day-0 support from major cloud platforms and inference engines.

Confirmed

  • Parameters and architecture: 2.4T total parameters, 95B active, 512 built-in experts, combining full attention with linear attention (Gated DeltaNet and Gated Attention); an FP8 version is available.
  • Capabilities: officially supports 1M context, with tunable reasoning and parallel tool calling, designed for coding and long-horizon agent workflows.
  • Multimodal: developer @multimodalart confirmed native vision and agent tool-calling support from the Hugging Face page config; Together AI describes it as Qwen's first multimodal model above 1T parameters.
  • License: per @cedricchee, MIT-like but with a commercial revenue threshold for large AI infrastructure or coding-assistant businesses; lenient for most startups.
  • Ecosystem: available on Together AI, Baseten and DeepInfra; Day-0 support from SGLang and Nebius Token Factory; DeepInfra pricing as low as $2/M tokens.
  • Real-world test: the SGLang team showed the model autonomously running for 4.5 hours to optimize its own SGLang inference serving and verify performance.

Unconfirmed

  • Context length descriptions vary: official claims 1M, Together AI lists 256K, DeepInfra lists 262K; no unified official clarification yet.

Why it matters

  • This is the first time Qwen has opened Max-tier flagship weights, and the 2.4T scale makes it one of the most closely watched ultra-large open-weight models.
  • SGLang's test shows the model can handle hours-long autonomous engineering tasks, marking a breakthrough in long-horizon agentic capability for open models.

1 more related posts →

Episode 2 · Unsloth Shrinks Qwen3.8 by 91% for Local Deployment (2026-08-13, 3 posts)

Unsloth AI introduced Dynamic 1-bit quantization, successfully compressing the massive Qwen3.8-2.4T-A95B model from 4.9TB to 397GB, reducing its size by 91% while maintaining impressive performance. The team also released a local deployment guide to lower hardware barriers.

Episode 3 · vLLM Announces Day-0 Support for Qwen3.8 2.4T Model (2026-08-13, 2 posts)

vLLM has announced Day-0 support for Alibaba's massive Qwen3.8-2.4T-A95B open-source model, providing out-of-the-box 4-bit quantization to enable single-node execution.

Episode 4 · Together AI Launches Serverless Inference for Qwen3.8-2.4T-A95B (2026-08-13, 2 posts)

Together AI has launched serverless inference support for the Qwen3.8-2.4T-A95B model, featuring 256K context and optimized for high-throughput coding and agentic workloads, enabling robust long-range task management.

Episode 5 · Alibaba's Qwen3.8-Max Lands on Five Hosted Platforms with Open Weights (2026-08-13, 6 posts)

Between August 13 and 15, Alibaba's flagship MoE model Qwen3.8-Max (also referred to as Qwen3.8-2.4T-A95B) rolled out across five hosted inference platforms — Modal, DigitalOcean Serverless, Nebius Token Factory, Fireworks and SiliconFlow — with open model weights. The model has 2.4T total parameters with 95B active, supports a full 1M-token context window, and targets autonomous agents and long-running coding tasks; SiliconFlow lists it at $2/M input and $6/M output tokens.

Confirmed

  • Model specs: 2.4T total parameters, 95B active, MoE architecture, full 1M-token context window
  • Modal (from Aug 13): deployed with a custom DFlash speculative decoder trained on tool-call-dense data
  • DigitalOcean Serverless: runs on NVIDIA HGX B300 GPUs, usage-based billing with no infrastructure management, designed for long-horizon coding
  • Nebius Token Factory: Day 0 partner launch, with open weights announced
  • Fireworks: Day-0 support as a launch partner, positioned for autonomous agents, heavy coding and long-context scenarios
  • SiliconFlow: model open-sourced and live; API priced at $2/M input and $6/M output tokens, with a discounted cached-input rate mentioned (post truncated); the official post highlights autonomous coding, deep research and end-to-end agent execution

Timeline

  • Aug 13: First launch on Modal (under the name Qwen3.8-2.4T-A95B)
  • Aug 14: DigitalOcean Serverless, Nebius Token Factory and Fireworks announcements follow
  • Aug 15: SiliconFlow launch with public API pricing

Why it matters

  • A flagship MoE model reached five major hosting platforms with open weights within three days, sharply lowering the barrier for long-context agentic coding workloads
  • Differentiated platform configurations (Modal's DFlash speculative decoding, DigitalOcean's B300 deployment) plus SiliconFlow's transparent per-token pricing give developers directly comparable deployment options