JOURNAL / 2026.10.22
Recursive releases Flint: examining local coding and automated research
Flint arrives with open weights and engines for coding agents. Its development with the Eureka Machine raises a more ambitious question: how to show that automating research actually improves the next model.
Recursive has introduced Flint, a coding model whose weights are released under Apache 2.0. The company attributes part of its development to Eureka Machine, its automated research system. One concrete example is searching for an architecture with fewer active parameters to accelerate local inference. That deserves attention, although moving from optimizing a model to demonstrating sustained recursive improvement still requires additional evidence.
What a project can use today
The model card describes a text model with 39.8 billion parameters, using approximately 2.5 billion per token. That selection reduces computation without making all the remaining weights dispensable. Its documented configuration accommodates 262,144 tokens, shared between input and output; it does not guarantee equal accuracy at every length.
There is a concrete route to using it. Ollama already lists it, with tools and reasoning, and shows a 23 GB file for the main variant. That download size should not be confused with the total memory required during a session. History, context and the runtime also consume resources.
For Apple Silicon, the rec engine repository requires an M4 Pro, M5 Pro or M5 Max and at least 48 GB of unified memory. It is optimized for a single session. Its comparison with llama.cpp measures generation on a 48 GB M4 Pro with four-bit weights and excludes initial prompt processing. That helps assess this particular setup, without promising identical speed on every laptop.
A server route is documented too. vLLM announced support on October 21, with a dedicated image while support is integrated into the project. Its optimizations target a Blackwell GPU; other configurations fall back to standard layers. Work on the runtime matters as much as parameter count when translating an efficiency claim into actual hardware performance.
Evaluations are not interchangeable either. The card warns that comparisons combine developer results obtained under different conditions. I would avoid turning the launch table into a universal ranking. To decide whether it works for a project, I would measure accepted changes, regressions and review time on its own tasks.
From automated experiments to recursive improvement
The account of Eureka has a verifiable history. In its June publication on automated research, Recursive describes a loop that proposes ideas, implements them, runs experiments and chooses the next step based on results. It presents fixed-budget training and kernel optimization tests, alongside public artifacts. It also acknowledges that assembling a good solution does not prove independent rediscovery of techniques: the underlying models may already know them.
That caution helps frame Flint. Producing a usable model allows outsiders to examine the result. Attributing how much it improved thanks to the research system additionally requires comparisons with equivalent budgets, recorded human interventions and failed experiments included. Demonstrating recursive improvement would require following several generations: checking whether the new model improves the researcher, and whether that researcher subsequently produces a better result under comparable conditions.
My reading is that Flint opens two distinct, valuable evaluations. A team can investigate whether it enables local work on code at an acceptable cost and quality. The community can ask for evidence of how its design decisions were discovered. The first question already has weights and engines to start with; the second needs a more complete experimental record than a success story. I would be particularly interested in seeing which proposals were rejected and which controls prevented optimization solely for the selected tests.
Primary sources consulted on October 22, 2026. We have not run Flint or reproduced its evaluations; performance claims come from their publishers.