AI Briefing
KOSign in

Domyn-Small-v1.0 support merged into llama.cpp

·2026.10.22 22:00

Key point

The 10B-parameter model uses roughly 2-4x fewer reasoning tokens than leading peers while maintaining comparable accuracy.

Details

  • Support for Domyn/Domyn-Small-v1.0 has been merged into llama.cpp, enabling local inference for the 10B-parameter open-weight model.
  • The model achieves reasoning accuracy comparable to leading 7-10B peers while spending roughly 2-4x fewer reasoning tokens, using approximately 32% of Qwen3.5-9B's and 35% of OLMo-3-7B-Think's token budget.
  • Features include dual-mode inference (thinking on/off), native 32k context extensible to 131k via YaRN, and first-class tool calling with 75.9 Non-Live and 68.3 Live scores on BFCL V3 single-turn benchmarks.
  • Developed by Domyn S.p.A., the model is initialized from Italia 10B, uses a 256,000-token SentencePiece BPE vocabulary, and is released under the MIT license.
inferencereasoningtool-callingopen-sourcellama.cppdomyn
Read originalShare on Threads

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

RedditAbliteratedOutTheBox enables live model abliteration in llama.cpp without fine-tuningHacker NewsDeepSeek open-sources DeepInfer inference framework for V4.1-FlashRedditNace.ai releases Kapitsa, an open-source Apple Silicon inference engine for Qwen and NemotronHuggingFace BlogIBM releases Granite Speech 5.0 with improved ASR accuracy and 1.6x faster Turbo CTCHuggingFace BlogNVIDIA Releases Nemotron 3.5 Embed, Rerank, and Parse Models for Agent RetrievalGeekNewsCopilotKit Releases OpenIntelligentUI, an Open-Source Alternative to ChatGPT Intelligent UI