Recursive Superintelligence, the startup chasing self-improving AI, published Flint on 21 October: a coding model with open weights under the Apache 2.0 licence. The artefact itself is straightforward to assess — compact, fast and free to download. The story wrapped around it is not. Recursive says its automated research system, the Eureka Machine, did nearly all of the work of building it, and nobody outside the company has verified that.
- Weights are open, hardware is not casual: Flint is a 39.8-billion-parameter mixture-of-experts model with roughly 2.5 billion active per token, but its practical builds require a machine with around 48 GB of unified memory.
- Decode speed is the selling point: Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro and above 150 on a 64 GB RTX Spark box — all company-measured.
- A narrow benchmark lead: Flint tops mixture-of-experts models under 50 billion parameters, yet loses to a dense 27B-class rival on four of six tests in Recursive’s own chart.
- Unverifiable provenance: there is no percentage, no audit and no external check behind the claim that AI built the model.
Architecture, context and the memory bill
Flint is a decoder-only mixture-of-experts Transformer: 49 layers (two dense followed by 47 MoE), 256 routed experts per MoE layer with six selected per token plus an always-active shared expert, and a split attention stack of 35 sliding-window layers and 14 full-attention layers. The documented context window is 262,144 tokens; the company blog mentions support up to 512,000 while advising deployments stay at or below 256,000, and the published GGUF repositories ship with the 262K figure. Stated knowledge cutoff: June 2026. Tool calling is OpenAI-compatible, with launch support across llama.cpp, Ollama, LM Studio, vLLM and SGLang.
The “runs on a laptop” framing survives only with small print. The Apple Silicon GGUF build weighs 26 GB and its card requires an Apple Silicon Mac with at least 48 GB of unified memory; the NVIDIA variant is 22.9 GB and recommends 48 GB of GPU or unified memory on Spark-class hardware. The main BF16 checkpoint runs to roughly 80 GB and asks for a B300 or H200 per replica, and Transformers modelling code is still missing.
Benchmarks: a qualified lead and an awkward panel
Recursive’s published numbers are 54.9% on Terminal-Bench 2.1, 74.2% on SWE-bench Verified and 89.7% on SWE-bench Pro v2, with 80.3 on LiveCodeBench v6, 89.2 on IFBench and 95.1 on IFEval. Against the peer Qwen3.6-35B-A3B it leads clearly on agentic and instruction-following rows (Terminal-Bench 54.9 versus 44.9; SWE-bench Pro v2 89.7 versus 73.8) with LiveCodeBench effectively tied. But a dense rival in the 27–31 billion range, Qwen 3.8, beats it outright in four of six tests, including Terminal-Bench (73.0 to 54.9), SWE-bench Verified (75.6 to 74.2) and LiveCodeBench (90.3 to 80.3). Two caveats come from Recursive itself: rival scores were not re-run, and the card warns that “evaluation setups vary across sources”. The higher score on the harder Pro panel is also counterintuitive, and that panel is the one marked as internally measured.
What the Eureka Machine actually contributed
The concrete engineering is real and documented: architecture search that started from a 35-billion-total, 3-billion-active base and explored configurations; a quantisation recipe that pushes the output layer to 4 bits “against common practice” using nearly half the training steps of its baseline; and specialised kernels, including a Metal megakernel fusing the whole decode step on M4 and M5 chips. Chief executive Richard Socher, formerly Salesforce’s chief scientist, said on X that the low-bit work shrank Flint by 71% with almost no loss in quality, while Nvidia reported the kernels sped the model up 55% on RTX Spark machines. A small team of people and agents built it from scratch in under four months.
Socher told Wired that “nearly every aspect of the entire model pipeline was actually done and improved upon by AI”.
The official blog is more cautious, saying the system “helped develop” nearly every part of Flint’s model-building stack — and it gives no percentage or time saved. The only inspectable prior evidence of the system’s power is older: a June post claims it improved the best community result on a small-model training task and cut a two-year-old speed record from 79.7 to 77.5 seconds, with a footnote conceding the team was still awaiting hardware access, so the record was never homologated. That post also admits reward hacking proved “especially difficult” on one task.
A useful academic counterweight is a preprint reviewing 1,250 papers that separates bounded self-refinement, an established practice, from open-ended recursive self-improvement, which is constrained by reality grounding, collapse dynamics and compute, and whose gains hinge on verifier quality. Recursive’s work fits comfortably in the first category. It is also worth noting that the system shares its name with a book Socher published in September.
What to watch next
Adoption, not mythology, will settle this. Within a day the Hugging Face builds had drawn roughly 3,000 downloads across the two Apple Silicon variants plus 1,561 for the NVIDIA one, and Ollama already lists 16.5K downloads — while the Hacker News thread sat at 24 points with two comments, both about speed. Recursive dismisses self-improvement fears as “more science fiction than reality”, even as US lawmaker Ro Khanna has planned a bill to ban self-improving AI and Google’s Sergey Brin has said Gemini now helps build the next Gemini. Flint’s weights can be inspected by anyone; the narrative around them cannot, and that asymmetry is the real story of this launch.