Recursive releases Flint, an open-weight, 40B-parameter coding model developed with its Eureka Machine AI system, which automated most of the development
Recursive makes the case that the benefits of using AI to build AI outweigh the risks. — For the past few days I've been testing …
WiredWill Knight
Context & Ripple Effects
Recursive's release turns the broader push toward recursive self-improvement at AI labs into a concrete software-development claim: its Eureka Machine automated most of the work behind a coding model that users can access as open weights. Anthropic had already said Claude authored more than 80% of code merged into its codebase, making automated model development an increasingly visible competitive capability.
The launch also joins an agentic-coding market where xAI positioned Grok Code Fast 1 around speed and economy. Recursive's local-use framing, echoed by launch-day discussion from LM Studio and Recursive executives, makes deployment control and API-free use part of the product proposition rather than merely benchmark performance.
First-order effects
Developers and local-model platforms gain access to Flint's weights, while Recursive can use the release as evidence that Eureka Machine can automate a large share of model-development work.
Recursive's launch-day distribution through local tooling makes its coding model immediately relevant to users seeking to run agentic coding workloads on their own hardware.
Second-order effects
Agentic-coding rivals, including xAI, face competition on local performance and operating-cost positioning as well as model quality.
Hardware and local-inference software providers gain a clearer workload to optimize for: long-context coding agents that users want to run outside hosted API workflows.
Third-order effects
If automated research systems repeatedly produce deployable models, the limiting advantage for AI labs shifts from headcount alone toward the quality of their automated development loops and their ability to validate outputs.
Open-weight releases paired with increasingly automated development sharpen the governance trade-off: availability broadens independent deployment while oversight must keep pace with models whose internal development and reasoning can be harder to inspect, a concern raised around OpenAI's recurrent-depth approach.
The trend: AI labs are combining recursive development systems with open-weight, locally deployable models, making iteration speed and inference control parallel competitive fronts.
We're excited to bring you Flint by @Recursive_SI in LM Studio on day 0! • 40B MoE, 2.5B active • 262K token context window • Apache 2.0 • Trained for local agentic coding model Requires about 24GB to run locally on your machine, performs impressively on benchmarks.
We are releasing our first model, Flint, a best-in-class local OSS coding model that runs at 150+ tokens/s on an @nvidia DGX Spark and 125+ tokens/s on a MacBook Pro M4—developed from scratch in less than four months with help from @Recursive_SI's autoresearch system.
7/ My bet is that we'll use both: powerful agents in the cloud, and increasingly capable models right beside us. Flint is our first step toward that local future. Try it on something you're building: https://docs.recursive.com/
4/ My ideal version of that everyday experience is fast, customizable, private, and free. Flint is a step toward that: open weights, running on your hardware, without API fees. Over 125 tokens/sec on an M4 Pro MacBook Pro at short contexts.
5/ We built Flint from scratch in under four months with a small team and our automated research system. It leads MoE models under 50B total parameters on agentic coding benchmarks. The system helped develop everything from training recipes to inference kernels.
3/ You'll want agents that can take on enormous amounts of work and keep going for hours or days. You'll also want something fast beside you while you're making things. Try an idea, see what happens, change your mind, try again.
2/ At OpenAI, when we were building the first version of Codex, there was a debate: focus on the local, interactive experience, or build for long-running agents in the cloud? I always thought we'd need both.
1/ Introducing Flint. It's a pretty good model for agentic coding locally. We @Recursive_SI think a lot of everyday coding will move to models running on your laptop or a server down the hall. Here's why we built it.
Flint is also a real-world test of the Eureka Machine, our system for automated AI research. It runs the full research loop: ideation, implementation, and validation. Then it uses the results to decide what to try next. We used it in nearly every part of how Flint was built. This…
Flint is the best-performing model in its class (MoE models under 50B) on agentic coding benchmarks like SWE-bench Pro v2, Terminal-Bench 2.1, and SWE-bench Verified. It also rivals similarly sized dense models that are much slower to run.
Introducing Flint, @Recursive_SI's first model: a fast local coding model built with our automated research system, the Eureka Machine. • Best-in-class fast local coding • Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark • Built from scr…