Key Takeaways
- Recursive Flint is the first model from Recursive, Richard Socher's AI startup, announced on October 21, 2026 as a fast local coding model with open weights.
- It is a roughly 40B-parameter Mixture-of-Experts model (about 39.8B total) that activates only around 2.5B parameters per token, which keeps local inference fast.
- Recursive claims 125+ tokens per second on a MacBook Pro and 150+ on an RTX Spark, and says the model was built from scratch in under four months using its Eureka Machine research system.
- No public download page or official build has been published yet, so for now you can only watch the preview and wait for the open weights to appear; check the official Recursive site for updates.
Quick Facts
| Type | Open-weight local coding model (MoE) |
|---|---|
| Devices | MacBook Pro, RTX Spark, local inference engines |
| Price | Not announced; free trial mentioned in opencode |
| Languages | Not confirmed; check official site |
| Highlight | 125+ tokens/sec locally, 40B knowledge at 2.5B cost |
| Learning curve | Easy to try, moderate to self-host |
| Category | AI & Models |
| Creator | Richard Socher |
| Heat | 🔥 3.4M Heat |
| Last updated |
What Recursive Flint Actually Is
Recursive Flint is the first model released by Recursive, the AI startup led by Richard Socher. Announced on October 21, 2026, it is described as a fast local coding model built with Recursive's automated research system, the Eureka Machine. The company says it was built from scratch in under four months, which is unusually quick for a model of this size and suggests the automated pipeline did much of the heavy lifting. Flint is positioned as private, unmetered intelligence that runs on your own machine rather than in a remote data center.
Technically, Flint is a Mixture-of-Experts model with roughly 40 billion total parameters, reported at about 39.8B, but it activates only around 2.5B parameters per token. That design is the key to its speed: instead of running the whole network for every token, it routes each token through a small subset of experts. The result is a model that carries broad coding knowledge but costs far less compute per token, which is what makes local, high-speed inference realistic on a laptop or a consumer GPU.
The open-weight label matters. Open weights mean the trained parameters are published so anyone can download them and run the model through their own inference engine, subject to whatever license Recursive attaches. That is different from a closed API-only model, where you can only send requests to a company's servers. Recursive says the weights are available now, though as of this writing no public download page or official build has been published, so treat availability details as something to verify on the official site.
- 40B total, 2.5B active
The Mixture-of-Experts design gives Flint a large knowledge pool while keeping per-token compute low, which is why local speeds are plausible.
- Built by the Eureka Machine
Recursive says an automated research system handled much of the development, producing the model in under four months.
- Open weights, not just an API
You are meant to run Flint yourself on your own hardware, not only through a hosted service.
- Coding-focused
Flint is tuned for programming tasks such as code completion, refactoring and debugging rather than general chat.
How the Speed Claims Work and What They Mean
Recursive claims Flint reaches over 125 tokens per second on a MacBook Pro and 150+ tokens per second on an RTX Spark. Tokens per second is the standard measure of how fast a model generates text; for coding, higher throughput means completions appear almost as fast as you can read them, which changes how you interact with an assistant. At 125 tokens per second, a short function can be written in a couple of seconds, making the model feel closer to autocomplete than to a chat bot.
Those numbers depend heavily on hardware, quantization and the inference engine you use. Quantization shrinks the model's numbers to use less memory, which usually speeds things up at some cost to quality. Different engines also handle Mixture-of-Experts routing with different efficiency. So treat the headline figures as best-case results on the hardware Recursive tested, and expect your own throughput to vary. The honest way to judge Flint is to benchmark it on your machine with your own code.
The deeper point is the ratio between knowledge and compute. A dense 40B model would need to run all 40B parameters for every token, which is far slower and heavier. By activating only about 2.5B parameters per token, Flint aims to deliver the coding ability of a much larger model at the running cost of a small one. That trade-off is the central engineering idea behind the model, and it is why the community is paying attention.
- Tokens per second is the metric to watch
It measures generation speed directly; higher numbers mean less waiting while you code.
- Your results will differ
Quantization level, RAM, GPU and inference engine all change real-world throughput.
- MoE routing is the trick
Only a small slice of the network runs per token, keeping compute low while retaining a large parameter pool.
How to Try Flint and Run It Locally
Because no public build or official page has been published yet, the first step is simply to watch the preview and follow Recursive's official announcements. Recursive has said the open weights are available and that you can run them locally on your favorite framework or try the model for free in the opencode coding environment. Until a download link and license appear on the official site, do not assume you can install it today; check back regularly and read the license terms carefully when they are posted.
When the weights do become downloadable, running Flint locally follows the same pattern as other open-weight models. You pick an inference engine that supports Mixture-of-Experts architectures, download the model files, and load them with enough memory to hold the weights. A 40B model is large, so expect to need a machine with substantial RAM or VRAM, or a quantized version that fits your hardware. The engine handles tokenization, routing and sampling; you just point your editor or terminal at it.
For a lower-friction start, the opencode coding environment is the route Recursive highlighted for trying Flint for free. That lets you evaluate the model's coding behavior before committing disk space and setup time to a local install. A sensible workflow is to test Flint in opencode on real tasks from your own projects, note where it succeeds and fails, and only then decide whether to self-host it for private, unmetered daily use.
- Start with the preview and official site
No public build exists yet, so announcements and the official Recursive site are your source of truth.
- Try it free in opencode first
Testing in a hosted coding environment avoids a heavy local setup before you know if the model suits you.
- Plan for the memory footprint
A 40B model needs serious RAM or VRAM; a quantized build may be necessary on laptops.
- Read the license before shipping
Open weights still come with terms; check what commercial and redistribution use is allowed.
How Flint Compares and Who It Is For
Flint enters a crowded field of coding models, and its pitch is specific: fast local inference with open weights rather than top-of-leaderboard benchmark scores. Comparable open-weight coding models exist, but many are dense architectures that run more slowly on consumer hardware, or larger Mixture-of-Experts models that need data-center-class memory. Flint's claimed advantage is the combination of a 40B knowledge pool with roughly 2.5B active parameters, which targets the sweet spot of capable and runnable on a single machine.
The privacy and cost angle is the strongest argument for Flint. Running locally means your code never leaves your machine, and there is no per-token billing or rate limit once you have the weights, which Recursive frames as private, unmetered intelligence. For developers working on proprietary code, or for anyone tired of usage caps, that is a meaningful difference from API-only assistants. The trade-off is setup effort and hardware cost, plus the fact that a local model may lag the largest hosted models on the hardest reasoning tasks.
The right audience is developers who want a fast, private coding assistant they control, and tinkerers who enjoy running models locally. If you need the absolute strongest reasoning on complex, novel problems, a large hosted model may still win. If you want quick completions, refactors and explanations without sending code to a server, Flint is aimed squarely at you. Because no public build is out yet, the practical advice is to watch for the release, test it in opencode, and benchmark it against whatever assistant you use now.
- Privacy is the headline benefit
Local inference keeps your source code on your own hardware, with no per-token charges.
- Speed over leaderboard bragging rights
Flint optimizes for fast, usable completions rather than maximum benchmark scores.
- Hardware is the real cost
You trade subscription fees for the RAM or GPU needed to run a 40B model comfortably.
- Unconfirmed details remain
License terms, exact supported languages and final availability should be verified on the official site.
Step-by-Step Guide
Watch the preview and note the claims
Start with Recursive's short preview of Flint to understand what is being promised: a fast local coding model with open weights. Write down the specific claims, such as 125+ tokens per second on a MacBook Pro, so you can verify them later.
Check the official site for the release
No public build or download page has been published yet, so the official Recursive site is where the weights, license and documentation will appear. Check it periodically and read the license terms before planning any commercial use.
Try Flint free in opencode
Recursive says you can try the model for free in the opencode coding environment. Use it on real tasks from your own projects to judge completion quality, speed and how well it follows your instructions.
Set up local inference when weights are available
Once the weights are downloadable, load them in an inference engine that supports Mixture-of-Experts models. Make sure your machine has enough RAM or VRAM, or choose a quantized build that fits your hardware.
Benchmark against your current assistant
Measure tokens per second and, more importantly, how often Flint's suggestions are correct on your codebase. Compare it side by side with whatever coding assistant you use today before switching your daily workflow.
Frequently Asked Questions
What is Recursive Flint?
Recursive Flint is the first AI model from Recursive, the startup led by Richard Socher. It is an open-weight coding model with roughly 40 billion total parameters in a Mixture-of-Experts design that activates only about 2.5B parameters per token. Recursive says it was built in under four months using an automated research system called the Eureka Machine, and it is designed to run fast on local hardware.
Can I download and run Flint right now?
As of the announcement, no public build, download or official page has been published; the model has only been shown in a short preview. Recursive has said open weights are available and that you can try it for free in opencode, but you should verify current availability and license terms on the official site before planning an install.
How fast is Flint, and what hardware do I need?
Recursive claims over 125 tokens per second on a MacBook Pro and 150+ on an RTX Spark. Real-world speed depends on your RAM or VRAM, the quantization level and the inference engine you choose. A 40B model is large, so expect to need a capable machine or a quantized build to run it comfortably.
Is Flint free, and is it private?
Recursive describes Flint as private, unmetered intelligence on your own machine, and mentions a free trial in opencode. Pricing for any hosted option has not been announced. Running the open weights locally means your code stays on your hardware, but you should confirm the exact license and any commercial restrictions on the official site.
How does Flint compare with other coding models?
Flint's pitch is speed and local privacy rather than maximum benchmark scores. Its Mixture-of-Experts design aims to give 40B-level knowledge at roughly 2.5B compute cost, which should make it faster on consumer hardware than dense models of similar size. The largest hosted models may still reason better on hard problems, so benchmark Flint on your own tasks.
