Tian Yuandong's Recursive ships its first model, built in under 4 months with AI
田渊栋新公司首个模型亮相！AI全程参与、不到4个月从零造
B · Useful, limited Jiqizhixin relays the launch of Recursive's first open-weight model, built by a small team with its automated research system; the method steps are described and a few independent speed tests are added, but benchmarks and timelines are self-reported.
Summary based on the full text·Original: about 5 minutes to read
In short
A Jiqizhixin article (22 October 2026) on Flint-40B-A2B, the first model from Recursive, the company co-founded by former Meta researcher Tian Yuandong and announced in May 2026. Flint is an agentic coding model designed to run on local hardware, released with open weights under Apache 2.0. Recursive says its automated research system, Eureka Machine, took part in almost every step: architecture search, data, training recipe, quantisation, infrastructure and inference kernels. A small team built it from scratch in under four months.
Key numbers
- 40B / 2.5Btotal and active parameters per token of the Flint-40B-A2B mixture-of-experts model
- under 4 monthstime from scratch to release, with a few staff plus the Eureka Machine system
- 125+ tok/sgeneration speed on a MacBook Pro with M4 Pro at short context, per Recursive
- 0.7maximum benchmark point gap between the 4-bit NVFP4 model and the 16-bit model
- ~46 tok/sspeed a developer reported on an RTX 3060 with 12 GB memory at 256K context
As reported by the source.
Chinese·Web page·opens jiqizhixin.com
What it reports
- Architecture: mixture of experts with 40B total and 2.5B active parameters, chosen by Eureka Machine from a 35B/3B baseline after searching expert configurations and widths.
- Benchmarks (self-reported): best among MoE models under 50B total on SWE-Bench Pro v2 and Terminal-Bench 2.1; top scores on IFBench and IFEval; some dense models still score higher on coding.
- Speed: 125+ tokens a second on an M4 Pro MacBook Pro and 150+ on an Nvidia RTX Spark with 64 GB; more than 5x faster than dense 25B-40B models with similar coding scores.
- Independent reports: 114-132 tokens a second on an M5 Max with 64 GB; about 46 tokens a second on an RTX 3060 with 12 GB at 256K context.
- Quantisation: the output layer was also quantised to 4 bits; an offline distillation step from the 16-bit model halved training steps versus a quantisation-aware baseline, and the NVFP4 model stays within 0.7 points of the 16-bit model.
- Kernels: custom low-bit, MoE and attention kernels, plus one fused Metal kernel for the whole decode pass on M4 and M5 chips.
- Context window up to 512K tokens, with 256K recommended. Recursive advises splitting complex tasks and giving more guidance than with large closed models.
How they got it
The article draws on Recursive's launch blog and earlier description of Eureka Machine, plus speed tests posted by developers. No third-party benchmark evaluation or ablation of Eureka Machine's contribution is shown.
Limits
- Benchmark results and the under-four-month timeline are self-reported.
- The share of work done by Eureka Machine versus staff is not quantified.
- Comparisons are limited to models below 50B total parameters.
Why it matters for companies
A capable coding model that runs on a laptop at usable speed makes offline, data-local coding assistance a realistic option for firms with strict data rules, at no API cost. The bigger signal is the claim that an automated research loop did most of the engineering for a competitive model in months; if it holds, model development cycles shorten further and small teams can compete on narrow model classes. Treat the benchmarks as provisional until independent evaluations appear.
Key passages in translation
Chinese original
在少量员工和 AI 智能体的共同参与下，从零开始，仅用不到 4 个月就完成了这款模型。
With a small number of staff working together with AI agents, the model was completed from scratch in under 4 months.
Chinese original
最终得到的 NVFP4 模型，在各项基准测试中的得分与 16 位模型相差不超过 0.7 分。
The resulting NVFP4 model scores within 0.7 points of the 16-bit model on every benchmark.
Chinese original
与大家熟悉的大型闭源编程模型相比，使用它时可能需要提供更多引导。
Compared with the large closed coding models people know, it may need more guidance in use.
Summary based on the full text. Read the original for the full argument and data.Read the original on jiqizhixin.com
- Tags
- Local inference