# Recursive launches Flint, a 40B open-weight coding model built by AI for local laptop inference

The Apache 2.0 mixture-of-experts model generates over 130 tokens per second on Apple Silicon while rivaling larger models on SWE-Bench Verified.

- Canonical: https://reviewblip.com/news/recursive-launches-flint-a-40b-open-weight-coding-model-built-by-ai-for-local-laptop-inference/
- Published: 2026-10-22T16:07:59.605Z
- Updated: 2026-10-22T16:07:59.605Z
- Author: ReviewBlip
- Category: Computing & Developer AI
- Type: News

![Technical benchmark graph displaying quality and decode speed of the Flint model on Apple Silicon M4 Pro across six agentic coding benchmarks](https://reviewblip.com/editorial-images/146b54b14a288f183ee35356c8fb656f54c6f0ad01caaf5c0ef6cc1c383b72ba)

Image credit: Recursive · M4 Pro inference speed and benchmark evaluation report

AI research lab Recursive, founded by former Salesforce Chief Scientist Richard Socher, released Flint on Wednesday evening—a 39.8 billion parameter mixture-of-experts (MoE) coding model released under an Apache 2.0 open-weight license. Engineered specifically for high-speed local inference on consumer hardware, Flint activates just 2.5 billion parameters per token, allowing software engineers to run fully offline coding agents on developer laptops without incurring cloud API costs or exposing proprietary source code to third-party servers.\[3\]

In benchmark tests conducted on an Apple MacBook Pro equipped with an M4 Pro processor and 48GB of unified memory, Flint achieved decode speeds of 131 tokens per second using optimized Apple MLX kernels—more than double the throughput of similarly sized dense models such as Google Gemma 4 and NVIDIA Nemotron 3.5 Lightning. On dedicated desktop hardware like NVIDIA's RTX Spark, inference throughput surpassed 150 tokens per second across context windows extending up to 512,000 tokens.\[3\]

Flint’s release serves as a real-world proof of concept for Recursive’s 'Eureka Machine,' an automated AI research pipeline that orchestrated nearly every stage of the model's development. Operating with a lean engineering team in under four months, the automated research system formulated model architecture pruning, synthesized targeted synthetic training corpora, executed low-bit quantization that compressed model footprint by 71% without performance degradation, and compiled custom inference kernels. On SWE-Bench Verified, Flint outperformed all open-weight models under 50 billion parameters.\[3\]

Our take: The release of Flint marks a critical inflection point for local developer tooling. Cloud-hosted coding assistants have dominated software engineering, but enterprise data governance, compliance mandates, and cumulative API token fees create substantial friction for daily development. A 40B MoE model that runs at 130 tokens per second entirely in local laptop RAM while solving real-world SWE-Bench bug fixes proves that edge-first developer AI is not just viable, but rapidly becoming the preferred modality for day-to-day programming.

## Sources
- [\[3\] TechCrunch / Recursive: Recursive launches Flint, a laptop coding model it says was largely built by AI](https://techcrunch.com/2026/10/21/recursive-launches-flint-a-laptop-coding-model-it-says-was-largely-built-by-ai/)