What happened
Recursive released Flint, a local mixture-of-experts (MoE) model with 40 billion total parameters that activates only 2.5 billion per token. The company states the model was built from scratch in under four months with the assistance of its automated research system, Eureka Machine. Flint is optimized for local inference, supporting up to 512k tokens of context and achieving speeds exceeding 125 tokens per second on M4 Pro MacBook Pro hardware and 150 tokens per second on NVIDIA RTX Spark systems. The model weights are available under the Apache 2.0 license, and Recursive is simultaneously releasing 'rec,' a specialized inference engine tuned for Apple M4 and M5 hardware.
Recursive has released Flint, a local mixture-of-experts model designed for fast agentic coding. The model contains 40 billion total parameters but activates only 2.5 billion per token, a design choice intended to cut computation required at inference. According to the source, a small team of people and agents built the model from scratch in under four months, utilizing Recursive’s automated research system, Eureka Machine, to assist in the process.
The model is optimized for local hardware, with Recursive reporting speeds of over 125 tokens per second on an M4 Pro MacBook Pro and over 150 tokens per second on a 64GB NVIDIA RTX Spark at short contexts. Flint supports up to 512k tokens of context, though the company recommends staying within 256k for optimal performance. The weights are released under the Apache 2.0 license and are compatible with popular local inference frameworks including llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang.
Recursive also released 'rec,' an inference engine specifically tuned for Flint on Apple M4 Pro and M5 Pro or Max hardware. This engine includes specialized kernels for low-bit projections, MoE operations, attention, and prompt processing, including a Metal megakernel that fuses the decode forward pass on M4 and M5 chips. The model uses NVFP4 quantization, which Recursive claims scores within 0.7 points of the 16-bit version across their internal benchmarks while nearly halving training steps against a quantization-aware baseline.
Source details: testingcatalog.com ↗
Why it matters
Flint addresses a critical gap in local AI capabilities by offering high-performance coding assistance without reliance on cloud APIs. By activating only a fraction of its parameters per token, it reduces computational overhead, making advanced agentic coding feasible on consumer-grade hardware. This shift allows developers to maintain data privacy and avoid recurring API costs, while the open-source license and broad framework support (llama.cpp, Ollama, vLLM) lower the barrier to entry for local deployment. The use of an automated research system to design the model also signals a potential shift in how AI architectures are developed, moving from manual tuning to automated optimization.
Flint targets developers who require coding agents that operate offline, keep data on-device, and avoid API fees. By leveraging a sparse MoE architecture, the model achieves performance comparable to dense coding models with 25 to 40 billion parameters while running more than five times faster, according to Recursive. This efficiency makes high-quality agentic coding accessible on consumer hardware that previously could not handle such workloads effectively.
The development process itself is notable for the role of Eureka Machine, which proposed ideas, implemented them, and tested results for the model's architecture, data, optimizers, and training recipes. This automated approach moved the design from a 35-billion-parameter baseline to the final 40-billion, 2.5-billion-active configuration, suggesting a new paradigm for AI model development that relies on automated search and optimization rather than purely manual engineering.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
What to watch next
Independent verification of the claimed SWE-Bench and Terminal-Bench scores is necessary, as these are currently self-reported by Recursive. Users should monitor the stability of the 'rec' inference engine on non-Apple hardware, as the source specifically highlights optimization for M4/M5 chips. Additionally, the practical usability of the 512k context window in real-world agentic workflows, particularly regarding latency and memory consumption, will determine the model's adoption among professional developers.
The performance claims, including leading MoE models below 50 billion parameters on SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified, are currently self-reported by Recursive. Independent third-party benchmarks will be required to confirm these rankings and the claimed speed advantages over comparable models.
While the model supports up to 512k tokens of context, Recursive advises users to stay within 256k. The practical impact of this limitation on long-running agentic tasks, such as refactoring large codebases, remains to be seen in real-world usage. Additionally, the specific performance of the 'rec' inference engine on hardware other than Apple M4/M5 chips is not detailed in the source, which may limit its immediate utility for users with different local setups.