# Recursive releases Flint, a free coding model with 2.5 billion active parameters that it says its own AI research system mostly built

The 39.8-billion-parameter mixture-of-experts model runs at about 131 tokens per second on an M4 Pro MacBook Pro, but every benchmark score Recursive has published comes from its own test runs.

- Published: 2026-10-26T06:26:37.653Z
- Canonical: https://polylog.news/ai/2026-10-26/recursive-releases-flint-a-free-coding-model-with-2-5-billio
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Runtime Wire](https://runtimewire.com/article/recursive-flint-local-coding-model-eureka-machine), [AI Weekly](https://aiweekly.co/alerts/recursive-opens-flint-a-40b25b-moe-coder-hitting-897-on-swe-bench-pro-v2-under), [MindStudio](https://www.mindstudio.ai/blog/recursive-flint-benchmark-coding-scores), [LM Studio](https://lmstudio.ai/models/recursivesi/flint), [The New Stack](https://thenewstack.io/meta-muse-spark-open/), [LLM Gateway](https://llmgateway.io/timeline)

Recursive released its first model, Flint, on October 21, [according to Runtime Wire](https://runtimewire.com/article/recursive-flint-local-coding-model-eureka-machine). The lab is run by Richard Socher, formerly Salesforce's chief scientist. Flint is a 39.8-billion-parameter mixture-of-experts model that activates about 2.5 billion parameters per token. It is built for coding agents and supports reasoning, multi-turn tool use and a 262,000-token context window, [according to its LM Studio listing](https://lmstudio.ai/models/recursivesi/flint). The weights are on Hugging Face under the Apache 2.0 license.

Recursive reports 54.9% on Terminal-Bench 2.1, 74.2% on SWE-bench Verified and 89.7% on SWE-bench Pro v2, [AI Weekly reported](https://aiweekly.co/alerts/recursive-opens-flint-a-40b25b-moe-coder-hitting-897-on-swe-bench-pro-v2-under). The Pro v2 figure needs the most caution. [MindStudio's review](https://www.mindstudio.ai/blog/recursive-flint-benchmark-coding-scores) notes that SWE-bench Pro v2 is a variant Recursive measured itself, that it is not comparable with public SWE-bench Pro numbers, and that no third party had reproduced any Flint score at launch. SWE-bench Pro was built to be harder than Verified, so a Pro-variant result 15 points above the Verified result indicates the two numbers are not measuring the same task set.

The more practical claim is speed. Recursive measured about 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip. At that rate a coding agent can run tool calls locally with no per-token fees, and the code never leaves the machine.

Recursive's larger claim concerns how the model was built. The company says its automated research system, the Eureka Machine, "helped develop nearly every part of Flint's model-building stack, from data and training recipes to infrastructure and inference optimization," and that "our small team of people and agents built it from scratch in under four months." Recursive published no technical paper, so there is no way to separate what the system did from what its human researchers did. The Russian-language channel [AI ML Big Data](https://t.me/ai_machinelearning_big_data/11204) notes that Recursive came out of stealth in May with $650 million raised before it had any model, based largely on the reputation of its founder.

Measured against other open-weight releases, Flint is not a frontier model. On Terminal-Bench 2.1, Meta's Muse Spark Open scores 76.4 and Moonshot's Kimi K3 scores 85.0, [The New Stack reported](https://thenewstack.io/meta-muse-spark-open/), and Muse Spark Open activates 42 billion parameters per token. What Flint offers is a high score for its active size. It arrived in the same week as Alibaba's Qwen3.6 27B on October 21 and Moonshot's Kimi K3.1 Pro on October 23, whose weights Moonshot has promised by November 4, [according to LLM Gateway's release timeline](https://llmgateway.io/timeline).

## What this means

The mechanism is cost. A model with 2.5 billion active parameters running locally can do the routine agent coding work (lint fixes, test scaffolding, small refactors) that developers currently pay for per token. The cheapest paid coding tiers at Anthropic, OpenAI and the coding-editor vendors built on top of them are the most exposed, and Apple's high-memory laptops are the hardware that gains. Capable small open-weight models are already an expected trend. What is new is the score-to-active-size ratio and the assertion that agents built most of the training stack, a claim that underpins a $650 million raise and has no published evidence behind it. The precedent is DeepSeek's R1 in January 2025: efficiency claims moved Nvidia shares sharply, and later independent work confirmed some claims and revised others. The likelier outcome is that independent harness runs within weeks land Flint's SWE-bench Verified score within a few points of 74.2% while the Pro v2 figure drops out of circulation, conditional on the released weights behaving the same outside Recursive's own setup. The read is wrong if public-harness results come in far below the reported number, which would make the launch a marketing exercise. Until Recursive publishes a technical report, the Eureka Machine remains a fundraising narrative rather than a verified result.

## What to watch

- Independent Flint runs on the public Terminal-Bench and SWE-bench Verified harnesses. Scores near Recursive's would confirm that small-active-parameter models can carry everyday agent coding, while a wide gap would mark the launch as a showcase for investors.
- Whether Recursive publishes a technical report separating the Eureka Machine's contribution from its human researchers', since that claim is what the company's valuation rests on.
- Whether Anthropic, OpenAI or the coding-tool vendors reprice their small fast tiers in coming weeks, which would signal that local open-weight models are taking routine agent traffic.