Recursive Launches Flint: Free 40B Coding Model That Runs on a Laptop

Recursive Launches Flint: Free 40B Coding Model That Runs on a Laptop

Recursive, the AI startup led by former Salesforce chief scientist Richard Socher, announced on Wednesday the official launch of Flint, its first model: a free coding model that runs locally on a laptop and is available for immediate download on Hugging Face under the Apache 2.0 license. The model uses a mixture-of-experts architecture with 40 billion total parameters, activating only 2.5 billion per task, and was built with the help of the automated Eureka Machine system in under four months.

Executive Overview

Recursive's Flint represents a significant shift in the economics of AI-assisted coding: a 40-billion-parameter model that activates just 2.5 billion parameters per task, runs entirely on a laptop, and carries no token fees. Available now on Hugging Face under Apache 2.0, it delivers 131 tokens/sec on a MacBook Pro M4 Pro—roughly double the speed of comparable models—and was built in under four months with the automated Eureka Machine system. The launch positions Recursive, backed by $650 million in funding at a $4.65 billion valuation, as a direct challenger to cloud-based coding services.

Deep-Dive Features & Architecture

Flint relies on a mixture-of-experts architecture that activates a small slice of the model for each task: 2.5 billion active parameters out of 40 billion total. This makes it far lighter than open models aimed at Mac hardware, such as Meta Muse Spark Open, which uses 816 billion parameters and requires at least a terabyte of memory. This design is what allows the model to run on an ordinary laptop while preserving code privacy, enabling offline use, and eliminating the token fees charged by cloud coding services.

The automated research platform Eureka Machine played a role in nearly every stage of building Flint, from pre-training to post-training and inference, according to the official launch video. Its contributions include designing a more sparse model structure, low-bit training that reduced Flint's size by 71% with almost negligible quality loss, and custom compute kernels that accelerate core operations. This enabled a small team to build the model from scratch in under four months.

Recursive was founded last year under CEO Richard Socher, former chief scientist at Salesforce, and emerged from stealth in May with $650 million in funding at a $4.65 billion valuation from investors including GV, Nvidia, and AMD Ventures. In June the company signed a $410 million compute agreement with Amazon Web Services, describing it as "less tied to headcount and more tied to the number of agents."

Benchmark & Competitive Performance

In Recursive's tests, Flint generated about 131 tokens/sec on a MacBook Pro with an Apple M4 Pro chip—nearly double the speed of comparable models that recorded 58 to 69 tokens/sec on the same device, including Google Gemma 4 and Nvidia Nemotron 3.5 Lightning. On the Nvidia RTX Spark platform it exceeded 150 tokens/sec. Flint also outperformed on SWE-Bench Verified among models under 50 billion parameters, according to the company.

• Speed: 131 tokens/sec on MacBook Pro M4 Pro; over 150 tokens/sec on Nvidia RTX Spark
• Efficiency: 2.5B active parameters from 40B total via MoE
• Footprint: 71% size reduction from low-bit training with almost no quality loss
• Privacy: fully local operation with no cloud dependency

Industry Impact & Enterprise Adoption

Flint's local-first design could reshape enterprise adoption of coding assistants by removing per-token costs and keeping proprietary code on-device. For organizations with strict data-sovereignty requirements, a free Apache 2.0 model that runs offline on a laptop offers a compelling alternative to cloud APIs. Recursive's $410 million AWS compute agreement and its investor roster—GV, Nvidia, and AMD Ventures—signal strong backing for scaling both model development and enterprise distribution. The model's availability on Hugging Face lowers the barrier for developers worldwide, including in the Arab region, to experiment locally without fees.

Conclusion

Recursive's Flint demonstrates that capable coding models no longer require massive cloud infrastructure. With 40 billion total parameters, 2.5 billion active, 131 tokens/sec on a MacBook Pro M4 Pro, and a free Apache 2.0 license on Hugging Face, Flint sets a new benchmark for local AI coding. Its rapid development via Eureka Machine and Recursive's $650 million funding underscore a broader industry shift toward efficient, private, and accessible AI tooling.

Media Source: TechCrunch AI | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis