AI Briefing
KOSign in

Flint model added to llama.cpp with 39.8B total parameters

·2026.10.22 06:29

Key point

The 39.8B parameter Recursive model features 47 MoE layers and a 262,144-token context window.

Details

  • Recursive's Flint model (FlintForCausalLM) is now available in llama.cpp following the merge of Pull Request #30928.
  • The model is a decoder-only mixture-of-experts Transformer with 39.8B total parameters and approximately 2.5B active parameters per token.
  • Architecture consists of 49 layers: 2 dense layers followed by 47 MoE layers, with 256 routed experts per MoE layer (6 selected per token) plus an always-active shared expert.
  • Attention mechanism uses 35 sliding-window layers (513-token window) and 14 full-attention layers, with 32 query heads and 4 key/value heads.
  • Documented context configuration supports 262,144 tokens, including prompt and generated output.
  • Vocabulary size is 150,144 tokens with a stated knowledge cutoff of June 2026.
inferencemodel-releasellama.cppmixture-of-expertsrecursiveflint
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

RedditFLE Studio: A Local Chat and Lab Interface for llama.cpp and FastLocalExpertsHacker NewsKrea 3 Launches: Frontier Image Editing Model Claims 10x Speed of Nano Banana ProRedditAbliteratedOutTheBox enables live model abliteration in llama.cpp without fine-tuningHuggingFace BlogNVIDIA Kumo Relational Outperforms Supervised ML on Relational Data Without Task-Specific TrainingNVIDIA BlogNVIDIA Showcases Local AI Hardware; Recursive Superintelligence Releases Flint for RTX SparkTLDR AISparkLLM Team Releases Spark-X2.6-4B, a 4B On-Device Model Matching 9B Performance