Open-source sparse model Flint: 40B params, 2.5B active, 125+ tok/s on M4 Pro for local agentic coding
agihouse_org · x · 2026-11-12
agihouse released Flint, the first model trained by its auto-research system (Eureka Machine): a sparse 40B-parameter model with only 2.5B active, purpose-built for local agentic coding.
Key points:
- Runs at 125+ tokens/sec on an M4 Pro MacBook Pro, targeting on-device deployment
- Ships a custom Metal megakernel for Apple M4/M5 chips, generated by the auto-research system
- Open weights, published on Hugging Face (RecursiveSI/flint), with a chat template supporting tool calls and reasoning content
It's an interesting take on combining sparsity, on-device kernels, and agent workflows for local coding assistants.
More from Infra
- Budget GPU server rental tips: OVH KS-6/7, FiberState, and asking for unlisted stock — realmrfakename · 2026-11-12
- AMD EPYC review image looks uncannily like logit artifacts from ViT segmentation models — giffmana · 2026-11-12
- llama-modchip runs 284B DeepSeek-V4-Flash IQ1 on a pure 64GB CPU with zero GPUs — pmttyji · 2026-11-12
- Agent Substrate Calculator Estimates Session Resume Speed Under Load — pjausovec · 2026-11-12
- Dell expands AI Data Platform: agents have a data-context problem, not a model problem — DavidLinthicum · 2026-11-12
- SGLang Megakernel Pushes Muse Spark to 1,109 tok/s, 2.1x Over Baseline — hsu_byron · 2026-11-12