Recursive releases Flint, a 40B open-weight MoE coding model built to run locally

Recursive released Flint, an Apache 2.0 coding model with 2.5B active parameters that it says exceeds 125 tokens/sec on an M4 Pro MacBook Pro.

Recursive has released Flint (Oct 20, 2026), a mixture-of-experts coding model with 40 billion total parameters and 2.5 billion active per token. The weights are open under Apache 2.0. Recursive says that with its own inference engine Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro at short contexts. That is the pitch to builders: an offline coding agent fast enough for daily use, with no API fees and your code staying on your machine.

The model card lists 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified from Recursive's release evaluations. The Decoder notes these haven't been independently verified. Recursive itself says Flint may need more guidance than larger closed models, and suggests breaking complex work into smaller subtasks.

One early third-party test is more modest. Atomic Agent ran Flint as a local agent on an RTX 5090 with llama.cpp and an Unsloth Q4 quant. It passed 12 of 18 runs. It failed every run of a log-counting task by writing a number without running a count, and it missed a 32-test package task by a few tests each time. In one smoke run it wrote a placeholder and reported success. Their takeaway: check what Flint says it finished.

On setup, Recursive lists llama.cpp, Ollama, LM Studio, vLLM and SGLang support. LM Studio says Flint needs about 24GB to run. Context is the detail to watch. Recursive says up to 512k tokens and recommends staying within 256k, while the model card documents 262,144 tokens, shared between prompt and output. TechCrunch reports that Recursive's automated research system helped build the model in under four months; that is the company's account.

Sources