# News: Recursive ships Flint, a 40B open coding model that runs on a laptop

> Recursive Superintelligence released its first model, Flint, on 21 October 2026: a mixture-of-experts coding model with roughly 40 billion total parameters and 2.5 billion active per token, published on Hugging Face under the Apache 2.0 licence. The company reports 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1, and generation above 125 tokens per second on a MacBook Pro with an M4 Pro chip (above 150 on an RTX Spark laptop). Recursive says its automated research system, the Eureka Machine, built nearly every part of the model pipeline in under four months with human oversight. The benchmarks are company-reported; TechCrunch and The Next Web both note the scores are not independently verified, and the model card warns Flint may need more task decomposition than larger closed coding models.

Human page: https://testedactually.com/news/recursive-flint-local-coding-model/
Mirror: https://testedactually.com/news/recursive-flint-local-coding-model.md
Published: 2026-10-23

## What changed

Recursive Superintelligence released its first model, Flint, on 21 October 2026: a mixture-of-experts coding model with roughly 40 billion total parameters and 2.5 billion active per token, published on Hugging Face under the Apache 2.0 licence. The company reports 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1, and generation above 125 tokens per second on a MacBook Pro with an M4 Pro chip (above 150 on an RTX Spark laptop). Recursive says its automated research system, the Eureka Machine, built nearly every part of the model pipeline in under four months with human oversight. The benchmarks are company-reported; TechCrunch and The Next Web both note the scores are not independently verified, and the model card warns Flint may need more task decomposition than larger closed coding models.

## What it means for buyers

Laptop-class local coding models now have a new open-weight entry that claims frontier-adjacent coding scores at consumer-hardware speeds, which strengthens the free local lane our Ollama and LM Studio profiles serve and gives buyers of per-seat coding assistants a local fallback to trial. Treat the published scores as vendor claims to re-test rather than findings, and note the local lane's usual caveats apply: no managed updates, and quality depends on the host tooling around the model.

## Sources

- [1] Introducing Recursive Flint (Recursive announcement) - https://recursive.com/updates/introducing-recursive-flint (company announcement, model specs and Eureka Machine framing, 21 October 2026, checked 2026-10-23)
- [2] Recursive launches Flint, a laptop coding model it says was largely built by AI (TechCrunch) - https://techcrunch.com/2026/10/21/recursive-launches-flint-a-laptop-coding-model-it-says-was-largely-built-by-ai (independent report, benchmark claims and funding background, 21 October 2026, checked 2026-10-23)
- [3] RecursiveSI/flint model weights (Hugging Face) - https://huggingface.co/RecursiveSI/flint (Apache 2.0 weight release, checked 2026-10-23)

TestedActually tracks pricing and programme changes as dated, sourced events: https://testedactually.com/news/.