Flint targets agentic coding with optimizations for Apple and Nvidia hardware
According to Recursive, Flint is optimized for fast local agentic coding on Apple and Nvidia hardware. It is available for free under the Apache 2.0 license; the published performance test results come from the company.
Flint supports agentic coding on local hardware and, according to Recursive, uses optimizations for Apple chips and Nvidia RTX Spark devices. The company reports speeds of over 125 tokens per second with a short context on a MacBook Pro with an M4 Pro chip and over 150 tokens per second on Nvidia RTX Spark with 64 GB of memory. The published benchmark results have not yet been independently verified.
Recursive made the model available for free on Hugging Face under the Apache 2.0 license, which permits use and modification. Its mixture-of-experts architecture has a total of 40 billion parameters, activating 2.5 billion per token. The model supports a context of up to 512 000 tokens, but the company recommends a maximum of 256 000.
The model can run without an internet connection using llama.cpp, Ollama, LM Studio, vLLM and SGLang. The company also released its own tool, rec, for Apple Silicon. The author of the Wired article tried the model on a Mac with 48 GB of RAM to create several small programs and gave it a favorable assessment. The company acknowledges that complex tasks require more guidance than with large proprietary models and recommends breaking them down into smaller steps.
According to Recursive, the Eureka Machine system assisted in almost every stage of development, including architecture design, training and computational optimization. Development took place under human supervision. The release of Flint does not in itself demonstrate full recursive self-improvement, in which the resulting model would demonstrably improve the ability of the system to develop further models.
Why it matters
The availability of a model that runs locally expands the options for coding without an internet connection and without sending code to external servers. Computer hardware and task complexity remain practical constraints; the published speed with a short context does not indicate performance at the maximum context length.
Release card
Flint
Recursive
- Specifications
- The model has a total of 40 billion parameters, of which it activates 2.5 billion per token.
- Context
- The model supports up to 512 000 tokens, but Recursive recommends a maximum of 256 000 tokens.
- Inputs
- The model generates text, including program code.
- Licence
- Apache 2.0
- Price
- Free.
- Availability
- The weights are available to download for free on Hugging Face. The model can run locally using llama.cpp, Ollama, LM Studio, vLLM and SGLang. Recursive also offers its own tool, rec, for Apple Silicon.
- Rychlost generování na MacBook Pro s čipem M4 Pro Více než 125 tokenů za sekundu According to measurements by Recursive, the model achieves this speed with a short context, but the results have not yet been independently verified.
- Rychlost generování na Nvidia RTX Spark Více než 150 tokenů za sekundu Recursive reports this speed on a device with 64 GB of memory, but the results have not yet been independently verified.
- Enables local agentic coding without sending code to external servers.
- Can be used to create small programs and simple websites.
- According to Recursive, complex tasks require more guidance than with large proprietary coding models.
Recursive compares the model with dense models with 25 to 40 billion parameters and claims it runs more than five times faster. In coding benchmarks, it also compares the model with mixture-of-experts models with fewer than 50 billion parameters; the results have not yet been independently verified.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
What was added since the original report
Verified updates
-
Flint is optimized to run locally at high speed on Apple and Nvidia hardware.; Flint is also intended for agentic coding.
- Flint is optimized to run locally at high speed on Apple and Nvidia hardware.
- Flint is also intended for agentic coding.
-
Flint is available to download for free on Hugging Face.; The model is released under the Apache 2.0 license.; The model has 40 billion parameters, of which it activates 2.5 billion.; The model can run locally without an internet connection.; A small team developed the model from scratch in less than four months.
- Flint is available to download for free on Hugging Face.
- The model is released under the Apache 2.0 license.
- The model has 40 billion parameters, of which it activates 2.5 billion.
- The model can run locally without an internet connection.
- A small team developed the model from scratch in less than four months.
Two audiences, two different impacts
What this means
For individuals
Developers can use Flint as an offline assistant for smaller coding tasks. For larger assignments, the company recommends planning to break up the work and provide ongoing guidance to the model.
For a business
Running the model locally allows a company to process source code on its own equipment and avoid cloud service fees based on the volume of text processed. However, the economic benefit depends on the available hardware and the amount of human guidance required.
DevelopmentCheck the original
Event sources
confirmed by 2 independent sources · 2 publishers, 2 independent. We count feeds from the same owner only once.