Recursive, co-founded by Yuandong Tian, open-sources Flint, a 40B local coding model

On October 22, Zhidongxi reported that last night, Recursive Superintelligence (hereinafter referred to as Recursive), an AI startup co-founded by former Meta FAIR research director Yuandong Tian, launched its first local coding model Flint. The model focuses on high-speed inference, achieving a generation speed of over 150 tokens/second on NVIDIA RTX Spark and over 125 tokens/second on a MacBook Pro equipped with an Apple M4 Pro chip, up to 2 times that of comparable models.

▲Recursive officially announces the launch of the Flint model, image source: X

Yuandong Tian subsequently reposted the news on Platform X and introduced four highlights of Flint.

▲Yuandong Tian posts about Flint's debut, image source: X

In terms of performance, Flint achieved scores of 54.9%, 89.7%, and 74.2% on the three agent coding benchmarks Terminal-Bench 2.1, SWE-Bench Pro v2, and SWE-Bench Verified, respectively, outperforming other MoE models with total parameter counts below 50 billion.

▲Flint's performance on Terminal-Bench 2.1, SWE-Bench Pro v2, and SWE-Bench Verified, image source: Recursive

In terms of training, Flint was built entirely from scratch, going through tokenizer development, pre-training, mid-training, supervised fine-tuning (SFT), and reinforcement learning (RL). In terms of inference speed, Flint's maximum generation speed can reach up to 2 times that of comparable models.

It is worth mentioning that Recursive took only 4 months to complete the setup of the entire Flint model training pipeline.

Some netizens said after trying Flint that the model could complete the tasks they gave it well, but the execution time was longer than expected. In addition, the netizen also reported that even with thinking mode turned off, the model's answers were still too verbose.

Another netizen said that what attracted him most was that Flint activates only 2.5 billion parameters per token. He is currently running about 50 agents on an M5 Pro device equipped with 48GB of memory, and the system often terminates some processes early due to memory pressure, while compute power has not yet become the bottleneck. Therefore, he hopes Flint can reside in local memory, be called at any time, take over some daily tasks originally handled by Qwen, and is considering adding it to his own model scheduling system.

Flint's debut has also brought renewed attention to the AI startup behind it, Recursive.

In May this year, Recursive officially made its public debut and announced that it had secured $650 million in funding (about 4.368 billion yuan), with a company valuation of $4.65 billion (about 31.246 billion yuan).

At the time, when Yuandong Tian announced joining Recursive, he mentioned that he would take Recursive Self-Improvement (RSI) as the core research direction, hoping to enable AI systems to autonomously conduct experiments and continuously improve themselves, thereby advancing the automation of scientific research.

About 5 months later, Recursive delivered its first model, Flint. Recursive co-founder and CEO Richard Socher, known as the "father of prompt engineering," also said on social platform X that his team developed Flint from scratch with the help of the automated research system Eureka Machine.

Flint can run directly on users' own devices, helping developers analyze codebases, fix program bugs, implement new features, and write tests.

▲Richard Socher, co-founder and CEO of Recursive, image source: X

Currently, Flint has been open-sourced under the Apache 2.0 license. Developers can obtain the model weights through Hugging Face and deploy and run it locally with compatible inference tools.

Open-source addresses:

GitHub:

Hugging Face:

1. Leading coding benchmark performance, challenging MoE models of the same scale

Flint adopts a decoder-only Mixture of Experts (MoE) architecture, with a total parameter count of 39.8 billion and about 2.5 billion parameters activated per token during processing.

Compared with dense models that need to invoke all model parameters for each inference, the MoE architecture can select some experts to participate in computation based on the input, thereby reducing the number of parameters that need to be activated for a single inference.

In the six benchmark tests Terminal-Bench 2.1, SWE-bench Pro v2, SWE-bench Verified, LiveCodeBench v6, IFBench, and IFEval, the BF16 version of Flint scored an average of 80.6 points, second only to Qwen3.8's 84.7 points and higher than Muse Glimmer's 74.8 points, Qwen3.6's 71.7 points, and Qwen3.5's 68.9 points.

Specifically, Qwen 3.8 leads Flint by 18.1 and 10 percentage points on Terminal-Bench 2.1 and LiveCodeBench v6, respectively; Flint, meanwhile, leads the other comparison models on the two instruction-following tests, IFBench and IFEval.

The NVFP4 quantized version of Flint scored an average of 80.0 points across the six tests, 0.6 points lower than the BF16 version; on the LiveCodeBench v6 test, its score was 0.5 percentage points higher than the BF16 version.

▲Flint's performance on coding benchmarks, image source: Recursive

Overall, Flint's coding performance is comparable to dense models with a total parameter count of 25 billion to 40 billion, while its generation speed is 5 times that of those models.

In terms of context length, Flint supports a context of up to about 512,000 tokens and recommends that developers keep actual usage within 256,000 tokens to achieve better coding results.

However, compared with closed-source coding models of larger parameter scale, Flint may require clearer instructions when handling complex tasks. Recursive recommends that developers break complex tasks into smaller subtasks to improve execution results.

II. Optimized for local device operation, inference speed up to over 150 tokens/sec

In terms of inference speed, test data published by Recursive shows that on an Apple M4 Pro device equipped with 48 GiB of memory, Flint 40B A2B using the MLX inference framework achieves a generation speed of 131.22 tokens/sec in short-context scenarios, about 2 times that of Qwen3.6 35B A3B (65.55 tokens/sec) and 2.3 times that of Nemotron 3.5 Lightning 30B A3B (58.00 tokens/sec), and also noticeably faster than Muse Glimmer 30B (14.54 tokens/sec).

On NVIDIA RTX Spark, the generation speed of the Flint NVFP4 version reaches 155.82 tokens/sec.

▲Flint inference speed published by Recursive, image source: Recursive

For different hardware, Recursive also optimized dedicated compute kernels. These optimizations increase Flint's generation speed on RTX Spark by 55% without modifying the model weights.

Recursive also launched its self-developed inference engine rec, used to run its language models locally.

Official tests show that on an Apple M4 Pro device equipped with 48 GB of memory, when using the 4-bit quantized model, rec's token generation speed reaches 1.4 to 2.6 times that of llama.cpp.

The test covered different input lengths from 1K to 252K tokens, generating 1024 tokens each time, with each group of tests repeated 5 times; the statistical results do not include the time spent on the prompt prefill stage.

▲Comparison of token generation throughput between the inference engine rec and llama.cpp on Apple M4 Pro, image source: GitHub

In terms of deployment, Flint supports local operation through tools such as SGLang, Ollama, and LM Studio. The model supports offline use, and developers can choose a suitable operating method based on their own hardware conditions to reduce dependence on remote services.

III. Automated system participates in R&D, optimizing model architecture and inference efficiency

During Flint's development, Recursive continuously improved model training and inference efficiency through continuous experimentation, evaluation, and iterative optimization.

The company's self-developed automated research system Eureka Machine can automatically propose research plans, write implementation code, and conduct experiments, then decide the next direction to try based on experimental results. The system can also explore multiple technical routes in parallel, use previous experimental results to adjust subsequent plans, and combine effective improvements from different plans.

With the help of the automated research system Eureka Machine, its team developed Flint from scratch in less than 4 months. During the R&D process, Eureka Machine also participated in training data processing, optimizer selection, and training plan design, and optimized the model infrastructure and inference performance.

Model architecture adjustment is one example.

The baseline plan Recursive initially tested had a total parameter count of 35 billion, activating 3 billion parameters per token. Subsequently, the automated research system Eureka Machine explored different expert configurations and network widths, ultimately selecting an architecture with a total parameter count of about 40 billion and about 2.5 billion parameters activated per token.

Recursive believes that among the configurations it has tested, this plan better balances operating speed and coding capability.

After model training was completed, the system continued to optimize the model's memory footprint and inference efficiency.

In the quantization stage, Eureka Machine attempted to compress the parameter precision of the model's output layer to 4 bits and added an offline distillation step to compensate for the performance loss that low-precision processing may cause. The final model using the NVFP4 format scored less than 0.7 points different from the 16-bit precision version on its benchmarks.

Recursive's model training time optimization experiment went through 6 rounds of iteration, with training time gradually decreasing from an initial 141.5 seconds to 77.5 seconds, an overall reduction of about 45%. The final round adopted a recursive optimization approach, further compressing training time to 77.5 seconds, 2.2 seconds faster than the previous record and setting a new low for this group of experiments.

▲Experimental results of Eureka Machine optimizing model training time, image source: Recursive

Recursive hopes to extend this R&D approach to subsequent model development, allowing the R&D results of each generation of models to in turn improve Eureka Machine's research capabilities, thereby accelerating the development of the next generation of models.

Recursive emerged from stealth in May 2026, with offices in San Francisco, USA, and London, UK. Its research areas involve automated algorithms, model training, and programming Agents capable of self-improvement.

The company's team already has more than 25 people, and many of its members have previously worked at enterprises and research institutions such as Salesforce, Uber, OpenAI, Google DeepMind, Google Brain, and Meta.

Recursive's co-founder team has a total of 8 people, including co-founder and CEO Richard Socher, Yuandong Tian, Tim Rocktäschel, Tim Shi, Josh Tobin, Caiming Xiong, Jeff Clune, and Alexey Dosovitskiy.

▲The eight co-founders of Recursive Superintelligence, image source: Recursive

Recursive co-founder and CEO Richard Socher is a well-known researcher and serial entrepreneur in the AI field.

He once studied under Stanford University professor and AI researcher Andrew Ng. His research areas include natural language processing and deep learning, and he has participated in related research such as ImageNet and GloVe. His Google Scholar citations exceed 240,000.

He is also one of the early researchers who promoted the application of neural networks to natural language processing, and is called the "father of prompt engineering" for his early systematic research on prompt engineering.

Socher once founded the AI startup MetaMind, which was acquired by enterprise software company Salesforce in 2016. After that, he founded the AI search company You.com and served as Salesforce's chief scientist and executive vice president.

It is worth mentioning that there are three Chinese members in Recursive's founding team.

Yuandong Tian previously served as director of research science at Meta's Fundamental AI Research team (FAIR), with research areas involving reinforcement learning, planning and efficiency, as well as theoretical research on large language models.

He received his bachelor's and master's degrees from Shanghai Jiao Tong University and obtained his doctorate from the Robotics Institute at Carnegie Mellon University in the United States in 2013. After graduating, Yuandong Tian worked as a software engineer on Google's self-driving car team from 2013 to 2014, then joined Meta FAIR to conduct AI research for nearly ten years, and left Meta in 2025.

▲Recursive co-founder Yuandong Tian, image source: Yuandong Tian's personal website

Caiming Xiong previously served as senior vice president of AI research at enterprise software company Salesforce and has long been engaged in AI research and model development. Before joining Salesforce, he worked as a senior researcher at MetaMind and had experience working with Socher.

He received his bachelor's and master's degrees from Huazhong University of Science and Technology, later obtained a doctorate in computer science and engineering from the University at Buffalo, State University of New York, and completed postdoctoral research in statistics at the University of California, Los Angeles.

▲Recursive co-founder Caiming Xiong, image source: Caiming Xiong's personal website

Tim Shi is an alumnus of Tsinghua University's Yao Class and once pursued a doctorate at the Stanford University Artificial Intelligence Laboratory, with research areas involving machine learning and AI systems.

In 2017, he joined OpenAI and participated in early AI Agent research. Subsequently, he co-founded the enterprise AI company Cresta and served as co-founder and chief technology officer, promoting the application of AI technology in enterprise sales and customer service scenarios. After that, he participated in founding Recursive Superintelligence.

▲Recursive co-founder Tim Shi, image source: Tim Shi's personal website

The other co-founders also have different research and entrepreneurial experiences. Tim Rocktäschel is a professor at University College London and has conducted AI research at Google DeepMind; Jeff Clune has long studied open-ended algorithms and self-improving AI.

Conclusion: Both speed and practicality, local coding model development accelerates

The launch of Flint reflects the emphasis of local coding models on runtime efficiency and deployment experience beyond code generation capability.

Running the model locally reduces reliance on cloud computing power, helps lower the need to transmit code and data to remote servers, and gives developers more flexible ways to use it.

Recursive combines the mixture-of-experts architecture with the automated research and development system Eureka Machine, which both optimizes the model's local runtime efficiency and explores a new path of accelerating model R&D through automated systems.

Source: Recursive