# Recursive releases Flint coding model, credits AI research system > Recursive released Flint coding-model weights and a local engine on October 21. The files establish access; its speed, benchmark and AI-built claims have distinct evidence limits. By BIGCHANGE.AI Editorial Published: 2026-10-23T17:56:38.514Z Updated: 2026-10-23T17:56:38.514Z Canonical: https://bigchange.ai/blog/recursive-flint-local-coding-model-weights ![Conceptual charcoal illustration of a silver laptop viewed from a low side angle, with a dark screen bearing indistinct marks and a power cable connected at the right.](https://bigchange.ai/api/media/file/flint-local-mac-coding-model-hero-v1.png) AI-generated conceptual illustration by BIG CHANGE; no Flint interface, local run or measured result is depicted. Recursive released Flint on October 21 with downloadable coding-model weights and a local inference engine. The [model repository](https://huggingface.co/RecursiveSI/flint/tree/main) contains the BF16 checkpoint, tokenizer, configuration, chat template and serving files under Apache 2.0. A separate [INT4 checkpoint](https://huggingface.co/RecursiveSI/flint-int4) and the [rec engine](https://github.com/recursive-oss/rec) give Mac users a smaller route to running it. Developers can inspect those artifacts; the speed and coding results remain company measurements. Flint puts a tangible result behind Recursive's claim that AI helped build a model. Its account describes an automated research system proposing and testing changes to architecture, quantization and inference code. The released artifacts establish the result of that work. Evidence that Flint improved the system that will design its successor has yet to be shown. ## The big change - **What changed:** Developers can download the weights for a coding model built with substantial reported AI research assistance and run a quantized version on supported local hardware. The available files make the model a usable artifact, while the development account remains Recursive's account. - **Why it matters:** A local coding agent can keep code and prompts on the operator's machine when its runtime and tools are configured that way. Flint's practical reach depends on memory, the chosen checkpoint and the agent harness, as well as model quality. - **What to watch:** Recursive's claim of a research process that improves through successive models calls for evidence about the next model and the process used to produce it. Comparable evaluations would need to separate gains from model design, compute, data and human decisions. ## What is available to run Flint is a mixture-of-experts model with 39.8 billion total parameters and about 2.5 billion active for each generated token, according to its [model card](https://huggingface.co/RecursiveSI/flint). Activating a subset helps limit computation during generation. It does not reduce the complete model to a 2.5-billion-parameter download. The BF16 repository lists about 79.6 GB of files, while Recursive's INT4 Mac checkpoint lists approximately 22.6 GB of weights. For the company's `rec` engine, the published requirements are an M4 Pro, M5 Pro or M5 Max Mac with at least 48 GB of unified memory and about 30 GB of free disk space. The engine is designed for one session and connects to pi, OpenCode or a local API. Other runtimes are described in the model card and launch post, but their memory requirements and performance cannot be inferred from `rec`'s specification. BIG CHANGE did not install a checkpoint or measure generation speed. The model card documents a 262,144-token context configuration, including input and output. Recursive's [launch post](https://recursive.com/updates/introducing-recursive-flint) says Flint supports up to 512,000 tokens but recommends staying within 256,000 for best results. Those statements describe different limits. The card also warns that a context capacity does not guarantee equal accuracy throughout it. ## What the speed and coding results measure Recursive reports more than 125 generated tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64 GB NVIDIA RTX Spark at **short contexts**, using its optimized inference stack. Its separate comparison chart describes average tokens per second over a 128,000-token session, excluding prompt prefill. The short-context rates cannot be presented as throughput across that long session. The `rec` repository reports a 1.4–2.6 times throughput advantage over llama.cpp on an M4 Pro with 48 GB, four-bit weights, prompts from 1,000 to 252,000 tokens, 1,024 output tokens and five runs per context; it also excludes prefill. These are Recursive's measurements, not tests by BIG CHANGE. On coding tasks, the model card reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. Flint's agentic scores average pass@1 across at least five attempts. Terminal-Bench used Terminus 2 with internet access but counted solutions obtained through retrieval as failures; the SWE-bench runs used mini-swe-agent without internet access. The card says comparison scores from other developers may use different setups. A rank drawn from such mixed protocols is less informative than the result for a specified task and harness. ## The AI-built claim has a narrower test Recursive says its Eureka Machine helped search model architectures, find a quantization recipe and optimize kernels for local hardware. The launch post gives concrete examples: a search from a 35-billion-total, 3-billion-active baseline to Flint's 40-billion-total, 2.5-billion-active design; an output-layer quantization and distillation step; and specialized kernels for Apple and NVIDIA hardware. These details describe AI-assisted development decisions as reported by the company. The public model files and engine establish the resulting artifacts, but do not provide an independent accounting of which proposals the system made, which people accepted or how much each contribution changed the outcome. Recursive calls Flint the starting point for a process in which stronger models make its research system better. That remains an ambition. As [BIG CHANGE's earlier survey coverage](https://bigchange.ai/blog/ai-self-improvement-successor-test) explains, improving one model or automating experiments differs from showing that a revised improvement method produces better successors across generations under comparable resources. Flint gives developers a model to evaluate and researchers a concrete case to scrutinize; it does not yet establish that stronger recursive claim. ## Sources & further reading - [Recursive, “Introducing Flint”](https://recursive.com/updates/introducing-recursive-flint) (release associated with October 21, 2026): The company's model, speed, benchmark and Eureka Machine claims, including its short-context speed statement and separate 128,000-token chart conditions. These are company measurements and descriptions, not independent replications. - [RecursiveSI/flint model card and files](https://huggingface.co/RecursiveSI/flint) (checked October 23, 2026): The BF16 files, Apache 2.0 license, architecture, documented context configuration, reported coding evaluations and their protocol caveats. Listing the files verifies publication, not successful operation on a reader's hardware. - [RecursiveSI/flint-int4 model card](https://huggingface.co/RecursiveSI/flint-int4) (checked October 23, 2026): The separate Mac-oriented quantized weights and approximate 22.6 GB size. That file size does not by itself specify the memory needed for a running agent. - [recursive-oss/rec repository](https://github.com/recursive-oss/rec) (checked October 23, 2026): The local engine's supported Mac hardware, memory and disk requirements, interfaces, one-session scope and stated comparison method. BIG CHANGE did not reproduce its speed results. - [vLLM's dated Flint integration record](https://github.com/vllm-project/vllm-project.github.io/blob/main/_posts/2026-10-21-flint.md) (October 21, 2026): A dated integration post coauthored with Recursive corroborates the release timing and documents a separate serving route. Its throughput claims depend on its stated GPU setup. ## Sources - [Recursive: Introducing Flint](https://recursive.com/updates/introducing-recursive-flint) — Official announcement; speed, benchmarks and AI research contributions are issuer claims; article body has no clear date. - [RecursiveSI/flint model card and files](https://huggingface.co/RecursiveSI/flint) — Published BF16 files, Apache 2.0 license, model specifications and company evaluation protocol. - [RecursiveSI/flint-int4](https://huggingface.co/RecursiveSI/flint-int4) — Official separate Mac quantized weights and approximate file size. - [recursive-oss/rec](https://github.com/recursive-oss/rec) — Official runtime repository, hardware requirements and issuer throughput method. - [vLLM Flint integration record](https://github.com/vllm-project/vllm-project.github.io/blob/main/_posts/2026-10-21-flint.md) — Dated integration record coauthored by Inferact and Recursive; corroborates release timing, not independent benchmarks.