Flint 40B by Recursive: Fast Local Coding AI Tested

Oct 21, 2026

Read: ~5m · You save: 3 min

Flint 40B by Recursive: Fast Local Coding AI Tested

Recursive's Flint 40B is a fast, open local coding AI. We run it on Ollama, test it on a doner kebab web page and a story, and review its benchmarks.

Overview

Recursive SI has released Flint, its first model. The company describes Flint as the fastest local coding model built with its automated research system, the Eureka machine. Flint is an open model designed for one function: serving as the reasoning core of a coding agent.

Model Specifications

Attribute	Detail
Architecture	Mixture of experts
Total parameters	About 40 billion
Active parameters per token	About 2.5 billion
Capabilities	Reasoning and tool calling across multiple turns
Context window	262K tokens
License	Apache 2.0
Released formats	GGUF in INT4, with additional quantizations on Hugging Face

The company's stated purpose is a simple workflow: explore a code base, fix a bug, and write a test.

Local Installation

The author ran Flint locally through Ollama. The installation process consisted of two steps:

Install the latest version of Ollama.
Pull the model.

Flint also runs with LM Studio, llama.cpp, vLLM, and SGLang.

During the test, the model used just over 32 GB of VRAM. Users can reduce VRAM consumption by lowering the KV cache size.

Running the full-precision version on a single GPU requires more than 80 GB of VRAM. The author notes that most users do not have hardware at this level.

Performance Tests

Coding Test: Doner Kebab Page

The author generated a web page without Hermes Agent or any other coding harness. The prompt produced a page showing a doner spit turning in front of a wall of gas flames. The model completed the page in 7 minutes.

The result includes:

Small blue flame bases beneath each flame
Dripping grease
A wisp of smoke
A button that displays a line of praise in 81 languages

The author assessed the output as follows:

The flames render well.
The visual fidelity is not high.
The skewer rotates.
The cone appears acceptable.
The meat texture is absent.

The author also asked the model to extract data from text and place it in a tab within a file. The model created the tab correctly. The HTML layout and CSS were rated as well structured.

The author states that Flint is a coding model, not a general language model. An attempt at a non-coding language task failed. The author advises against using Flint for scientific reasoning, creative writing, or similar tasks.

Creative Writing Test

The author asked the model to write a story about a child and a dinosaur who embark on an adventure. The model produced a story without failing. The author rates the result as average. For a 40-billion-parameter model, the author considers the result acceptable, given that the model does not position itself as a creative writer.

Benchmark Claims

The following results come from the model card and vendor benchmarks run on an Apple M4 Pro. The author notes that the benchmarks emphasize Apple hardware.

Quality and Speed

Flint sits at the edge of the quality-speed frontier.
Only Qwen 3.8 27B scores higher, and it runs at a fraction of Flint's speed.
Every other model in the chart trails Flint in quality or speed.
The chart includes Muse Glimmer, a recent model, while the remaining models are older.

Raw Speed

The MLX version of Flint exceeds 100 tokens per second at short context on the M4 Pro.
The llama.cpp version performs nearly as fast.
On the RTX Spark system, the NVFP4 version runs faster still.
Similarly sized sparse models run slower. Only the smallest dense models are faster.

Comparison with the Qwen Family and Other Models

Flint outperforms similarly sized sparse models on the agent coding test.
Flint shows particular strength in following instructions.
The larger dense Qwen 3.8 model still leads on the most difficult terminal and software engineering tasks.

Quantization

According to the chart, the NVFP4 version loses almost no accuracy compared with full precision. The vendor recommends the compact build.

Recursive Inference Runtime

Recursive also released a dedicated inference runtime, called Rec. Rec is tuned for Flint on Apple silicon. It fuses the entire decode pass into a single custom Metal kernel.

According to the company, Rec delivers 1.4 to 2.6 times more tokens per second than llama.cpp, including at very long context lengths. The author has not tested Rec because the author does not have access to an Apple machine. The author based the description on the GitHub repository and additional documentation.

Assessment

The author considers Flint impressive for coding tasks. The author reports that the model produced a working result quickly for the repository tested. The author states that Recursive should support additional hardware architectures beyond the Apple ecosystem.

Introduction to Flint

Recursive SI releases Flint, its first model, described as the fastest local coding model, built with the automated research system Eureka machine.

Flint is Recursive SI's first model
Described as the fastest local coding model
Built with their automated research system, the Eureka machine
Open model built for one job: being the brain of a coding agent

Model Architecture and Specifications

Flint is a mixture-of-experts model with about 40B total parameters, activating roughly 2.5B per token. It reasons, calls tools across multiple turns, has a 256K context window, and ships under Apache 2.0.

Mixture of experts, about 40 billion parameters in total
About 2.5 billion parameters active per token
Reasons and calls tools across multiple turns
256K context window (transcribed as '260 2K')
Released under Apache 2.0
GGUF in INT4 released; other quants available on Hugging Face

Local Installation and Running

The author runs Flint locally through Ollama, using about 32 GB of VRAM, which can be reduced by lowering the KV cache. Supported runtimes include LM Studio, llama.cpp, vLLM, and SGLang.

Installed via Ollama: install latest Ollama, then pull the model
VRAM use just over 32 GB while running
VRAM can be reduced by reducing KV cache
Also usable with LM Studio, llama.cpp, vLLM, and SGLang
Running unquantized requires more than 80 GB VRAM on a single GPU, which the speaker says most people do not have

Coding Test: Doner Kebab Page

Flint generated a doner kebab web page with flames, a rotating skewer, grease drips, smoke, and a 81-language praise button in about 7 minutes. The HTML layout and CSS looked good, though the meat texture was weak. Translation failed because Flint is not a language model.

Doner kebab page generated without Hermes Agent or any coding harness
Built in 7 minutes
Includes blue flame bases, dripping grease, smoke, and a button showing praise in 81 languages
Skewer rotates, cone looks okay, meat texture is weak
HTML layout and CSS were quite nice
Data extraction into a tab worked well
Translation failed; the speaker says Flint is not a language model and it is not for scientific reasoning or creative writing

Creative Writing Test: Kid and Dinosaur

A creative writing test about a kid and a dinosaur produced an average story. It did not fail outright, and the speaker considers it decent for a 40B coding-focused model.

Prompt: a story about a kid and a dinosaur on an adventure
Result: an average story, not a complete failure
Considered a good result for a 40B parameter model in this context format, though it could be better

Benchmark Charts: Quality vs Speed

The speaker reviews benchmark charts from the model card, measured on an Apple M4 Pro. Flint sits at the edge of the quality-speed frontier, with only Qwen 3.8 27B scoring higher, though at much slower speed.

Charts from the model card, run on Apple M4 Pro
Quality (higher is better) versus speed (further right is faster)
Flint is on the edge of the frontier
Only Qwen 3.8 27B scores higher, but runs at a fraction of the speed
Other compared models include Muse Glimmer and older models

Raw Speed Benchmarks

On an Apple M4 Pro, the MLX version of Flint passes 100 tokens per second at short context, with llama.cpp close behind. On an RTX Spark box, the NVFP4 version is faster. The NVFP4 version loses almost nothing compared with full precision.

MLX version exceeds 100 tokens per second on M4 Pro at short context
llama.cpp version is not far behind
NVFP4 version is faster on RTX Spark box
Similar-sized sparse models trail; only the tiniest dense models are quicker
NVFP4 loses almost nothing compared with full precision

Benchmark Against Qwen Family

Flint is compared with Qwen family models, Laguna, Muse Glimmer, and Java. It beats similar-sized sparse models on agent coding and is strong at instruction following, but the larger dense Qwen 3.8 wins the hardest terminal and software engineering tasks.

Compared against Qwen family, Laguna, Muse Glimmer, and Java
Beats similar-sized sparse models on agent coding test
Especially strong at following instructions
Qwen 3.8 dense still wins the hardest terminal and software engineering tasks
Flint is not best at everything

Rec Inference Runtime

Recursive also released Rec, its own inference runtime tuned for Flint on Apple silicon, which fuses the decode pass into one custom metal kernel. It claims 1.4 to 2.6 times more tokens per second than llama.cpp even at very long context. The speaker has not tested it.

Rec is Recursive's own inference runtime, tuned for Flint on Apple silicon
Fuses the whole decode pass into one custom metal kernel
Claims 1.4 to 2.6 times more tokens per second than llama.cpp, even at very long context
Speaker has not tested it because he has no Apple machine
Numbers taken from the GitHub repo and other docs

Conclusion and Opinion

The speaker finds Flint impressive for coding, especially its speed, but says Recursive should support other architectures rather than focusing on Apple.

Impressive results in coding, works very fast
Main criticism: should support architectures beyond Apple ecosystem