FUNDING · STARTUPS · #2369
Recursive unveils Flint, a 40B coding model developed using its 'Eureka Machine'
Recursive, a startup led by Richard Socher and co‑founded by Jeff Clune, is offering Flint, an open‑weight 40‑billion‑parameter coding model the company says was largely developed and optimized by its recursive AI system called the Eureka Machine. Recursive shared benchmark claims that Flint outperforms peers on coding tasks and has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures.
KEY POINTS
- Recursive, a startup led by Richard Socher and co‑founded by Jeff Clune, is offering Flint, an open‑weight 40‑billion‑parameter coding model the company says was largely developed and optimized by its recursive AI system called the Eureka Machine.
- Recursive shared benchmark claims that Flint outperforms peers on coding tasks and has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures.
- This is a notable demonstration of recursive self‑improvement tools being used to automate model design and optimization—potentially accelerating model development while raising questions about oversight and long‑term safety.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Recursive publicly released Flint, a 40‑billion‑parameter mixture‑of‑experts coding model (2.5B active parameters) that the company says was largely developed and optimized by an internal automated research system called the Eureka Machine. Flint is available as an open‑weight model on Hugging Face under Apache 2.0, designed to run locally on laptops and optimized for Apple and Nvidia hardware. Recursive published benchmark claims that Flint outperforms comparable models on coding tasks and reports fast local inference (e.g., ~125–131 tokens/sec on an M4 Pro, >150 on an RTX Spark). The company previously raised $650M at a $4.65B valuation from investors including GV, Nvidia, AMD Ventures, and Greycroft.
WHY NOW
The release is a concrete instance of using recursive/AI‑driven research (an AI that proposes, runs, and evaluates experiments) to design and optimize a production model quickly. That brings practical evidence into an ongoing debate: proponents point to faster, smaller‑scale model development and local deployment; critics and some researchers worry recursive self‑improvement techniques could reduce human oversight. The publication of performance claims and open release makes these debates and any verification efforts immediately actionable.
WHO IS AFFECTED
Directly affected parties include: developers and teams who may use local coding models (privacy/offline use), researchers and engineers comparing or benchmarking coding models, the AI safety and research community monitoring recursive‑improvement techniques, and investors and vendors involved in model infrastructure and hardware optimizations (Apple/Nvidia/RTX ecosystems are explicitly targeted).
CONFIRMED
• Recursive released Flint, a coding-focused language model. • Flint is a mixture‑of‑experts model with ~40 billion total parameters and ~2.5 billion active parameters per token. • Recursive says the Eureka Machine automated or assisted nearly every step of Flint’s development (proposing, running, evaluating experiments; contributing to training data, optimizers, architecture, inference tuning). • Flint is published on Hugging Face under the Apache 2.0 license and is intended to run locally on laptops, optimized for Apple and Nvidia hardware. • Recursive reports benchmark results (e.g., SWE‑Bench variants, Terminal‑Bench) claiming Flint outperforms comparable models under 50B on coding tasks and reports inference speeds ~125–131 tok/s on an M4 Pro and >150 tok/s on RTX Spark. • Recursive previously raised $650M at a $4.65B valuation from investors including GV, Nvidia, AMD Ventures, and Greycroft. • Recursive states the Eureka Machine and Flint were built in roughly four months by a small team with heavy use of the automated system.
UNCERTAIN
• Independent verification: Decoder explicitly notes benchmark results have not been independently verified; no third‑party benchmark reports are provided in the excerpts. • Degree of autonomy: sources vary in wording ("largely developed and optimized by AI"; "helped with nearly every step"; "proposed, ran, evaluated experiments"). The exact balance of automated vs. human decisions and specific human oversight mechanisms are not detailed in the excerpts. • Safety controls and stopping mechanisms: Recursive’s claim that the Eureka Machine is “incapable of escaping human control” and that humans could stop recursive processes is reported as company assertion; excerpts contain no technical description or third‑party assessment of safeguards. • Some technical details are incomplete in the excerpts (e.g., Decoder mentions a 4‑bit output quantization and a distillation step but the excerpt is truncated and lacks full data on impacts and tradeoffs). • Benchmark scope and comparators: company reports focus on coding benchmarks and comparisons with certain MoE and dense models under 50B, but full benchmark methodology, datasets, and independent baselines are not included in the excerpts.
WHAT TO WATCH
Observable signals to monitor (based on the supplied excerpts): - Independent third‑party benchmark/verification reports for SWE‑Bench, Terminal‑Bench, and other cited tests. - Release of detailed methodology, model cards, training data documentation, and safety/oversight descriptions from Recursive or third parties. - Community activity on Hugging Face: downloads, forks, community reproductions, and open‑source evaluations. - Follow‑up disclosures from Recursive on Eureka Machine autonomy, human‑in‑the‑loop controls, and technical safeguards. - Additional technical tests comparing Flint to other open models on diverse coding tasks and hardware performance measurements. - Regulatory or legislative mentions/actions concerning recursive self‑improvement techniques (the excerpts cite ongoing debate and proposed bills).
WHY IT MATTERS
This is a notable demonstration of recursive self‑improvement tools being used to automate model design and optimization—potentially accelerating model development while raising questions about oversight and long‑term safety.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
Recursive released Flint, a coding language model with a total of 40 billion parameters.
SUPPORTEDChecked 2026-10-22 · 3 supporting
WIRED refers to an AI coding model called Flint that the author tested, supporting the model's existence and name.
SUPPORTS TechCrunch AIMEDIA · 2026-10-22TechCrunch reports the public release of Flint and describes it as a 40B-parameter model.
SUPPORTS The DecoderMEDIA · 2026-10-22The Decoder describes Flint as a 40-billion-parameter model for programming tasks.
Flint uses a mixture-of-experts (MoE) architecture with about 2.5 billion active parameters per token.
SUPPORTEDChecked 2026-10-22 · 2 supporting
Recursive says Flint was largely developed and optimized by its internal AI system called the 'Eureka Machine'.
SUPPORTEDChecked 2026-10-22 · 3 supporting
WIRED quotes CEO Richard Socher saying the Eureka Machine automated and improved nearly every aspect of the model pipeline.
SUPPORTS TechCrunch AIMEDIA · 2026-10-22TechCrunch reports Recursive's claim that the Eureka Machine helped with nearly every step of Flint's development.
SUPPORTS The DecoderMEDIA · 2026-10-22The Decoder describes the Eureka Machine proposing, running, and evaluating experiments that contributed to Flint's development.
Recursive shared benchmark and speed results claiming Flint outperforms similar-sized models on coding tasks and runs faster on local hardware; those results are reported by the company and have not been independently verified in the excerpts.
SUPPORTEDChecked 2026-10-22 · 2 supporting
WIRED notes Recursive shared benchmark scores suggesting Flint outperforms peers on coding tasks, as reported by the company.
SUPPORTS TechCrunch AIMEDIA · 2026-10-22TechCrunch cites Recursive's reported token-per-second measurements and comparisons with other models on specific hardware.
CONTEXT The DecoderMEDIA · 2026-10-22The Decoder repeats the benchmark and speed claims but explicitly notes the results 'haven't been independently verified yet'.
SOURCES & TIMELINE
3For the past few days I’ve been testing an impressive open-weight AI coding model called Flint. It suggests that fully recursive self-improvement , or RSI—an exciting and controversial approach that uses AI to build AI—is getting closer. Flint is the first offering from Recursive, a company that aims to use RSI to advance both AI itself and other areas of science using AI . “Nearly every aspect of the entire model p…
Recursive , an AI startup founded last year, is starting to live up to its name. The company, which is trying to build AI that can improve itself, released its first model, Flint , on Wednesday. Flint is a free coding model that can run on a laptop and was built largely with the help of AI. Recursive’s automated research system, which it calls the Eureka Machine, helped with nearly every step of Flint’s development,…
The AI startup Recursive has released Flint, a language model for programming tasks that was developed largely by an internal AI system called "Eureka Machine." Flint uses a mixture-of-experts architecture with 40 billion parameters (2.5 billion active) and is optimized for fast local execution on Apple and Nvidia hardware. The model is freely available on Hugging Face under the Apache 2.0 license. Recursive's long…