AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
All posts Back News August 12, 2026 •4 minute read Qwen3.8-2.4T-A95B now available on Modal Harmya Bhatt Member of Technical Staff @racerfunction Gilford Ting Member of Technical Staff @gilfordting David Wang Member of Technical Staff @_dcw02 Richard Gong Member of Technical Staff @_gongy Will Hu Member of Technical Staff @_williamhu Greta Workman Product Marketing @gretaworkman Qwen3.8-2.4T-A95B just launched as an open weights model, and it’s now available on Modal. Over Qwen 3.7, the latest model sees substantial improvement across coding, work, research, and long-horizon tasks. We worked with Qwen ahead of the drop to bring day zero support to Modal Auto Endpoints, backed by SGLang and a custom DFlash speculator tuned to Qwen3.8’s shape. Try it out now as a Shared Endpoint. Speeding up inference with custom DFlash speculation Just getting the model running is one thing, making it fast is another. For this, we once again turn to a DFlash speculator model because—say it with us now—Speculation is all you need. A speculator only earns its keep when the target accepts the tokens it drafts, and acceptance comes down to whether the drafter has seen sequences like the ones it's predicting. Because of Max’s improvements in coding, research, and work (things that tend to use more tool calls) we leaned into that in our training data to increase accepts. Try it now Qwen3.8-2.4T-A95B (text only) is available for the next month as an OpenAI compatible Shared Endpoint with token-based pricing.