Flint
Recursive's Apache 2.0 40B mixture-of-experts coding model with 2.5B active parameters, built to run locally.
Overview
Recursive released Flint on October 21, 2026: a 40B-parameter open-weight coding model with 2.5B active parameters, 256 routed experts (6 per token), and a 262,144-token context window, under Apache 2.0. Recursive reports 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1 (versus 73.0% for dense Qwen 3.8 on its comparison), and over 125 tokens per second on a MacBook Pro with an M4 Pro; these are company-reported. 4-bit weights take about 23GB, targeting 48GB Apple Silicon or NVIDIA machines, and the company says most development was automated by its 'Eureka Machine' system.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 40B total / 2.5B active MoE, 256 routed experts, 6 selected per token
- 262,144-token context window
- Company-reported 74.2% SWE-bench Verified and 54.9% Terminal-Bench 2.1
- ~23GB at 4-bit; >125 tok/s claimed on M4 Pro MacBook Pro
- Runs via Ollama (flint:40b), LM Studio, llama.cpp, Unsloth, vLLM, and SGLang; OpenHands day-zero partner
Use cases
- Local, private coding agents on a 48GB laptop or workstation
- Open-model backend for agent harnesses such as OpenHands
Pricing
Free open weights under Apache 2.0; also listed with a free tier on OpenRouter.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 21 Oct 2026
- Record updated
- 23 Oct 2026
- Record last checked
- 23 Oct 2026