AI BriefingBasalt 0.4.0 releases Qwen 3.8 27B dense model with 776 tok/s structured output
·2026.11.16 02:30
Key point
The update adds multi-model routing for DeepSeek V4.1 Flash and video support alongside the new dense model.
Details
- Basalt 0.4.0 introduces the first dense model, Qwen 3.8 27B, achieving up to 776 tok/s on structured output and 262 tok/s in conversation.
- The quantized model maintains 92.8% top-1 accuracy compared to the full BF16 model with a KLD of 0.0262.
- Multi-model and router support is implemented, starting with DeepSeek V4.1 Flash, which runs at 90 tok/s conversation and 246 tok/s structured on a 5090 + 5060 Ti.
- New features include video support, files in chats, and up to 262K context length with prefill speeds of 12,000 tok/s at 4k context.
Technical Details
- Prefill performance is 12,000 tok/s at 4k context, slowing as context grows.
- Prose generation speed is 194 tok/s.
- The release includes fixes and improvements for chat, tools, thinking, images, and video.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
GitHub Next releases Ace, a local collaborative coding environment for humans and AI agents