Benchmark: Recursive's Flint 40B A2.5B Q4 achieves 128 tok/s on MacBook Pro M4 Pro 48GB using Kapitsa, an open-source Apple Silicon optimized engine, Claude/Codex/Pi/OpenCode compatible.
recursive released a model today and they claim new model that is optimized for coding on 32gb+ macs, its Flint, they claim to beat 35b sparse MoEs by Qwen and come close to qwen 27b dense. we went ahead and added support of Flint to Kapitsa, reaching SOTA decode/prefill on sub M5 chips.
Proof1 community thread
Key points
- try it (needs a Mac with 32 GB or more; the download is 22.6 GB):
- brew install nace-ai/tap/kapitsa kapitsa serve --model flint # OpenAI and Anthropic APIs on localhost kapitsa opencode --model flint # or claude, codex, pi kapitsa bench
- The model card's sampling: T 1.0, topp 0.95, topk 20.
- No speculative decoding anywhere: Flint has no draft model yet.
Sources (1)
- [1]Benchmark: Recursive's Flint 40B A2.5B Q4 achieves 128 tok/s on MacBook Pro M4 Pro 48GB using Kapitsa, an open-source Apple Silicon optimized engine, Claude/Codex/Pi/OpenCode compatible.r/LocalLLaMA (top, daily) · Oct 22, 12:17 PM
recursive released a model today and they claim new model that is optimized for coding on 32gb+ macs, its Flint, they claim to beat 35b sparse MoEs by Qwen and come close to qwen 27b dense. we went ahead and added support of Flint to Kapitsa, reaching SOTA decode/prefill on sub M5 chips.
try it (needs a Mac with 32 GB or more; the download is 22.6 GB):
Extractive summary: sentences quoted from the sources.
Before this
- Oct 21, 2026Recursive shows first result of its AI-driven research process with Flint
- Oct 21, 2026Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster
- Oct 20, 2026[AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
- Oct 20, 2026Do Diffusion LLMs Scale? Testing Mercury 3 + Decide
- Oct 20, 2026AnyLanguageModel 1.0: One Year of Local Models on Apple Platforms
- Oct 20, 2026MPS in PyTorch 2.15
