Recursive Ships Flint: a 39.8B MoE Coding Model With 2.5B Active Params and Its Own Inference Engine
BerryBlush_Rogers691 · reddit · 2026-10-22
Recursive released Flint, an Apache 2.0 coding-focused model (39.8B total / 2.5B active params, 256K context) supporting reasoning and multi-turn tool use, plus rec, a local inference engine integrated with pi and OpenCode that reportedly hits 1.4–2.6× the generation throughput of llama.cpp on an M4 Pro (prefill excluded).
Caveats: rec currently requires M4 Pro/M5 Pro/M5 Max with ≥48GB unified memory, is optimized for a single session, and official GGUF downloads remain large. The poster questions whether the throughput gain matters in full coding tasks once repo reading, tool results, and test runs are included, calling for a like-for-like rec vs llama.cpp benchmark on long conversations.
More from Infra
- PyTorchCon talk: QR, eigendecomposition and Cholesky kernels decide if better optimizers actually train faster — PyTorch · 2026-10-22
- 'This is my quant': CoreWeave seen stuck in a reflexive credit doom loop — vasuman · 2026-10-22
- Yaroslav Bulatov's bear case for local models: data centers win on intelligence per dollar — yaroslavvb · 2026-10-22
- vLLM brings native PyTorch serving to TPUs with open-source TorchTPU — PyTorch · 2026-10-22
- Wiz: stolen credentials from LiteLLM supply chain incident already abused for LLMJacking from two tracked IPs — wunderwuzzi23 · 2026-10-22
- Vantage Reserves Over 1GW of Heron Link Capacity for NVIDIA's 800 VDC AI Power Architecture — espricewright · 2026-10-22