Editorial illustration for Recursive's Flint local 40B MoE coding model tops 150 tokens/sec on RTX Spark
Recursive's Flint 40B MoE Model Hits 150 Tokens/Sec
Recursive Superintelligence put out its first model this week, and it's aimed squarely at people running coding agents on their own hardware instead of someone else's API. Flint is a mixture-of-experts model with 39.8 billion total parameters but only about 2.5 billion active per token, which is the whole point: it fits on a 48 GB Apple Silicon Mac or a single 48 GB Nvidia GPU, with 4-bit weights running in roughly 23 GB. The model card lists a 262,144-token context window, though Recursive's blog claims it stretches to 512k.
The benchmarks tell a split story. On IFEval, Flint hits 95.1%, and it scores 74.2% on SWE-bench Verified, numbers that put it ahead of other MoE models under 50B parameters on agentic coding tasks, according to Recursive's own chart. But Terminal-Bench 2.1 is where it struggles, landing at 54.9% against 73.0% for the dense Qwen 3.8.
Recursive says an internal automated research system called Eureka Machine built nearly every layer of the stack. The weights are on Hugging Face under an Apache 2.0 license, so anyone can inspect how that claim holds up against the architecture itself.
Recursive Superintelligence has released Flint, its first model. The Recursive Flint local coding model is a mixture-of-experts (MoE) model with 40B total and 2.5B active parameters. Recursive team states that its Eureka Machine, an automated research system, helped build nearly every part of the stack.
Why this matters
The speed numbers matter less than what they imply about hardware access. A 40B MoE model with only 2.5B active parameters hitting 150 tokens/sec on a 64GB RTX Spark, and 125+ on an M4 Pro MacBook, suggests local coding agents that don't need a data center lease to feel responsive. For developers weighing local inference against API calls to larger hosted models, Flint's active-parameter math is the pitch: you get something closer to a 40B model's knowledge with a fraction of the compute bill per token. The Apache 2.0 license means teams can actually inspect, fork, and fine-tune without licensing friction, which matters if you're building agent tooling you intend to ship commercially.
We'd still want independent benchmarks before trusting Recursive's own comparisons against "dense 25-40B models," since the company is grading its own homework here. The Eureka Machine claim, that an automated system built most of the stack, is worth watching too. If that holds up under scrutiny, it says as much about the tooling behind Flint as the model itself.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives 43% of Indian Enterprises Struggle to Control AI Costs: Splunk Report
From Archives Google Cloud Ships llm-d 1.0 With Cache-Aware TPU and GPU Routing
Common Questions Answered
What is the difference between Flint's total parameters and active parameters per token?
Flint has 39.8 billion total parameters but only approximately 2.5 billion active parameters per token. This mixture-of-experts architecture allows the model to maintain the knowledge capacity of a much larger model while requiring significantly less computational resources and memory during inference.
What hardware requirements does Recursive's Flint model need to run locally?
Flint can run on a single 48GB Apple Silicon Mac or a single 48GB Nvidia GPU with 4-bit weights using roughly 23GB of memory. This makes it accessible to developers who want to run coding agents on their own hardware without relying on cloud APIs or data center leases.
What throughput speeds does Flint achieve on different hardware platforms?
Flint reaches approximately 150 tokens per second on a 64GB RTX Spark GPU and 125+ tokens per second on an M4 Pro MacBook. These speeds demonstrate that local coding agents can achieve responsive performance comparable to cloud-based alternatives without requiring expensive infrastructure.
How did Recursive Superintelligence build the Flint model?
Recursive's Eureka Machine, an automated research system, helped build nearly every part of the Flint model stack. This automated approach enabled the company to develop an efficient mixture-of-experts architecture optimized for local deployment.
What context window size does the Flint model support?
Flint supports a 262,144-token context window, which provides substantial capacity for processing long coding tasks and maintaining conversation history. This large context window enables developers to work with extensive codebases and detailed project requirements without token limitations.
Further Reading
- Recursive Releases Flint, an Open-Weight 40B MoE Local Coding Model for Agents with 2.5B Active Params - MarkTechPost
- Bringing the Spark to Europe: Local Intelligence Powers Next-Generation AI Innovation - NVIDIA
- Introducing Flint - Recursive
- Recursive launches Flint: 40B MoE local coding model - AGI Hunt
