Recursive · 2026-10-21
Introducing Flint
“Our Eureka Machine helped develop nearly every part of Flint’s model-building stack, from data and training recipes to infrastructure and inference optimization.”
- Who
- Recursive
- Date
- 2026-10-21 The post shows no date; Recursive's updates page lists it as Oct 21, 2026.
- Evidence
- Not rated. A description of how the model was built. The benchmark figures measure Flint, not how much of the work the Eureka Machine did. how we rate evidence
- Replication
- Nothing to rerun. all claims
- Checked
- 2026-10-22, against the source page
- Links
- Source
Also on the page
- Recursive's stated goal: "to build recursively self-improving superintelligence to automate knowledge discovery."
- Recursive's bet for what comes next: "Our bet is that more capable models can make Eureka Machine a better researcher, helping us build the next generation faster and deliver more capability for the same cost."
- Architecture: "Starting from a baseline with 35 billion total parameters and 3 billion active parameters, our Eureka Machine explored different expert configurations and widths."
- Quantization: "Against common practice, it quantized the output layer (lm_head) to 4 bits." The offline distillation stage the Eureka Machine added "nearly halved the number of training steps compared with our quantization-aware training baseline."
- Kernels: "On M4 and M5 chips, it produced a Metal megakernel that fuses the entire decode forward pass."
- Recursive's earlier post, First Steps Toward Automated AI Research (listed on its updates page under Jun 11, 2026), "describes the system and its results on our initial test problems."
What got repeated wrong
- no primary sourceA model largely built by AI.The post says "Our small team of people and agents built it" and that the Eureka Machine "helped develop nearly every part" of the stack. It gives no figure for how much of the work AI did.