Recursive's Flint 40B: A Laptop Coding Model Built by AI
By @sharedot · · 8 pages
- AI Frontier
- Open Weight Models
- Local AI
- Agentic Coding
Recursive released Flint, a 39.8B-parameter open-weight coding model built largely by its Eureka Machine in under four months, under Apache 2.0.
What shipped: an open-weight MoE that runs on a laptop
Recursive SI released Flint on October 21, a mixture-of-experts coding model with 39.8 billion total parameters and roughly 2.5 billion active per token, published on Hugging Face as RecursiveSI/flint under Apache 2.0 with no revenue gate, no OEM clause and no service-operator trigger.
Weights ship as an ~80GB BF16 checkpoint plus quantized builds — BigGo Finance reports an NVFP4 version that Recursive says retains nearly all full-model performance. Vendor-reported throughput exceeds 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark.
Why it is surprising: AI built the model that builds models
The surprise is not just the size — it is the provenance.
According to BigGo Finance, Recursive's automated research system, the Eureka Machine, contributed to nearly every stage of Flint's development, from architecture search and quantization to low-level GPU kernel tuning, including a low-bit training recipe that cut model size by about 71% with minimal quality loss.
"Nearly every aspect of the entire model pipeline was actually done and improved upon by AI," founder Richard Socher said, framing the release as an early public test of recursive self-improvement — while acknowledging humans still set goals and can stop the process.
The evidence, and its footnotes
BigGo Finance reports Flint scored 89.7% on SWE-bench Pro v2, trailing only Qwen 3.8 in the company's comparisons, and OrcaRouter adds vendor-reported 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified — all vendor figures with no third-party measurement published yet.
OrcaRouter also flags an unresolved discrepancy: the announcement describes a 512,000-token context window with 256,000 recommended, while the Hugging Face model card documents 262,144.
Forkast News reports the evaluation protocol uses pass@1 over at least five attempts, counting web-retrieval solutions as failures to measure offline coding specifically.
Context: the four-month, $4.65B team behind it
Recursive, co-founded by former Salesforce chief scientist Richard Socher with Tian Yuandong and Jeff Clune, emerged from stealth with $650 million in funding at a $4.65 billion valuation, with backers including Google Ventures, Greycroft, Nvidia and AMD Ventures, BigGo Finance reports.
The company argues automating model-research trial and error lets a small team compete with far larger labs.
Forkast News notes Flint sits in the same unrestricted license tier as DeepSeek's DeepInfer engine and NVIDIA's Nemotron Retriever models — beyond the $20 million revenue gates four labs attach to their open-weight models.
The companion story: open-weight agents replicate MIT
The same day, TechBullion reported that Časlav Nedeljković of HAL8's AI Local Lab reproduced most central findings of MIT's SwarmWorld experiment using a quantized Qwen3.8-27B model on four consumer RTX 3090 cards — 50 agents, roughly 26 hours, 12,802 model calls, nine of twelve paired effects matching MIT's direction.
The collective advantage held and, notably, the smaller model benefited more from working in a society than it did alone.
The replication also surfaced a reproducibility issue: two 4-bit builds scored 0.41 versus 0.24 in MIT's validity pilot, suggesting KV-cache precision matters as much as model size.
What comes next
Both developments point the same direction: capability moving onto the developer's own hardware.
Forkast News frames Flint as the weights-to-watts argument at its smallest unit — no cloud dependency, no gate on growth, and if the frontier commoditizes down to local coding hardware, the watt in question sits on the developer's desk.
Open questions remain: Recursive has not disclosed Flint's build cost, OpenHands tested the model but has not independently validated its benchmark claims per BigGo Finance, and the SwarmWorld team plans KV-cache isolation tests, a 100-agent population run, and a fuller Vector settlement-gateway comparison.
Sources
- forkast.news › Recursive SI Launches Flint 40B: Open Weights That Run on the Developer's Own Machine
- finance.biggo.com › Recursive Debuts Flint: A Laptop-Sized Coding Model That AI Helped Build in Under Four Months
- orcarouter.ai › Recursive Flint 40B-A2.5B: The Fast Local Coding Model
- techbullion.com › MIT SwarmWorld Findings Replicated on Four GPUs Using an Open-Weight AI Model