The launch of Flint and the role of Eureka Machine
Recursive, founded by Richard Socher, presented its first model, Flint, on October 21, 2026. It is an open-weight language model designed specifically for agentic coding on local hardware. The most distinctive feature of this launch is not just the model itself, but the process that originated it: the company's automated research system, known as Eureka Machine.
According to the company, Eureka Machine intervened in almost every stage of development, from defining the architecture to optimizing inference. A small team of researchers and agents built the model from scratch in less than four months, a timeframe significantly shorter than usual in the industry for models of this scale. The company describes this process as a step toward its vision of recursively self-improving superintelligence, although for now the system operates under human supervision.
Technical architecture: 40B parameter MoE
Flint uses a Mixture-of-Experts (MoE) architecture with 39.8 billion total parameters, rounded to 40B in public communications. The key to its efficiency lies in activating only approximately 2.5 billion parameters per processed token. This configuration allows the model to maintain competitive reasoning capabilities while drastically reducing the computational load during text generation.
The model is designed as a transformer decoder with 49 layers, of which 47 are MoE-type and 2 are dense. It uses 256 routed experts, activating 6 per token plus one shared expert. The documented context window reaches 262,144 tokens, although the company recommends staying below 256,000 to optimize performance in coding tasks.
Flint technical specifications
Summary of the model's main characteristics according to the published technical sheet.
| Feature | Detail |
|---|---|
| Total parameters | 39.8 billion (40B) |
| Active parameters per token | ~2.5 billion |
| Architecture | MoE (47 MoE layers, 2 dense) |
| Max context | 262,144 tokens |
| License | Apache 2.0 |
Performance and inference speed
Recursive reports that Flint is up to twice as fast as comparable models in its class. On a MacBook Pro with an M4 Pro chip, the model reaches over 125 tokens per second, while on an NVIDIA RTX Spark with 64 GB of memory, it exceeds 150 tokens per second. These figures correspond to short contexts and use the company's optimized inference engine.
In terms of task-solving capability, the company cites 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified, and 54.9% on Terminal-Bench 2.1. It is important to note that these results are self-reported by Recursive and have not been validated by independent entities to date. The company acknowledges that the model may require more guidance and the division of complex tasks into subtasks compared to larger closed-source models.
Self-reported benchmarks
Performance results declared by Recursive at the launch of Flint.
| Benchmark | Score | Notes |
|---|---|---|
| SWE-bench Pro v2 | 89.7% | Self-reported |
| SWE-bench Verified | 74.2% | Pass@1, average of 5+ attempts |
| Terminal-Bench 2.1 | 54.9% | With internet access |
How the autonomous research worked
The Eureka Machine system did not just tune hyperparameters; it proposed and implemented structural improvements. According to the documentation, the system started from a reference configuration and explored different expert architectures to find the optimal balance between speed and accuracy. It also developed specialized kernels to accelerate low-precision operations and attention on specific hardware such as Apple Silicon and NVIDIA.
One of the most notable contributions was the optimization of quantization. Eureka Machine managed to compress the model to 4 bits in the output layer (lm_head) and applied an offline distillation stage that reduced the required training steps by almost half. The result was an NVFP4 version that remained within 0.7 points of the 16-bit model in internal tests, minimizing quality loss.
Eureka Machine workflow
Iterative process described by Recursive for the development of Flint.
- 01Proposal of ideas and alternative architectures
- 02Implementation of changes in data and training
- 03Execution of experiments and evaluation of results
- 04Optimization of kernels and inference for local hardware
- 05Iteration based on measured performance
Availability and hardware requirements
Flint's weights are available on Hugging Face under the Apache 2.0 license, allowing free use, modification, and distribution. The main checkpoint in BF16 format occupies approximately 80 GB, which requires high-performance hardware such as B300 or H200 GPUs for full execution.
To facilitate adoption on consumer devices, Recursive is preparing quantized variants for NVIDIA and Apple Silicon. These versions reduce the required memory, allowing the model to run on modern laptops. The model is compatible with popular frameworks such as llama.cpp, Ollama, LM Studio, vLLM, and SGLang, as well as a lightweight inference engine optimized for Apple devices.
Open questions
- There are no independent benchmarks validating the performance figures reported by Recursive.
- The exact percentage of the training process performed without direct human intervention has not been precisely specified.
- Quantized variants for consumer hardware are still in preparation and are not available in the main repository at the time of launch.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction