Qwen3.8-2.4T

	Qwen3.8-2.4T(huggingface.co)
	700 points by Philpax 9 days ago \| 168 comments
	tl;dr: Qwen3.8-2.4T-A95B is Alibaba's new flagship open-weight MoE model (2.4T total / 95B activated params, 92 layers, 512 experts) with a hybrid Gated DeltaNet + Gated Attention architecture, 262K native context extensible to ~1M tokens. It's a thinking-only text model with tunable reasoning_effort, targeting coding and long-horizon agentic tasks, and posts benchmarks competitive with Claude Opus 4.8 and GPT-5.6 on SWE-bench Pro, Terminal Bench, and PaperBench. Deployable via vLLM, SGLang, and TokenSpeed, with a hosted Qwen3.8-Max variant adding vision and built-in tools.
	HN Discussion: ~Model is impressively powerful but too large to serve without quantization work •Contextualizes release against competitors like Kimi k3 and DeepSeek V4 ↓Disappointment that open weights lack vision and full 1M context of hosted variant ↓Concerned about inaccessibility for local/consumer hardware users ↓Skepticism about the model's actual real-world performance versus benchmarks