Recursive's Flint Puts a Fast Coding Agent on Your Laptop

Recursive Superintelligence, Inc.—the company led by CEO Richard Socher—has released Flint, an open-weight mixture-of-experts model aimed at agentic coding on local hardware. Weights ship under Apache 2.0 on Hugging Face, alongside rec, Recursive's own inference engine tuned for Apple Silicon Macs. On October 21, 2026, the vLLM project added day-0 serving support. The pitch is familiar in 2026—run a capable coding agent without sending your repo to a cloud API—but Recursive is betting that a MoE architecture built from scratch, plus aggressive kernel work, can make that feel fast enough for daily use on a laptop or a compact NVIDIA desktop.

Key facts

What Recursive released

In its Introducing Flint post, Recursive frames Flint as a coding model for developers who want an agent that can investigate codebases, fix bugs, implement features, and write tests—with text in/out, reasoning, and multi-turn tool use. The company says a small team of “people and agents” built the model from scratch in under four months, using its Eureka Machine automated research stack for architecture search, quantization, and custom kernels.

Flint is not a thin fine-tune of an existing checkpoint. Recursive positions it as a MoE designed for the latency profile of local agents: fewer active parameters per token than a dense model of similar capability, with hardware-specific paths to keep decode fast. The same announcement pairs the model with rec on GitHub—a lightweight local inference engine that launches pre-configured pi (preferred) and OpenCode harnesses for day-to-day agent loops.

Key specs

How fast it is—and on what hardware

Recursive's speed claims assume its optimized stack at short contexts. With its optimized inference engine, the company says Flint exceeds 125 tokens/s on an M4 Pro MacBook Pro and 150 tokens/s on a 64GB NVIDIA RTX Spark, and can run up to twice as fast as comparable models in those setups. The rec repository claims 1.4–2.6× llama.cpp token throughput on M4 Pro; Recursive also cites a Metal megakernel that fuses the decode pass on M4/M5 silicon.

On the data-center side, Recursive documents a BF16 checkpoint served on one B300 or H200 GPU via an SGLang kit. vLLM's October 21, 2026 announcement adds day-0 Flint support—shipping via the vllm/vllm-openai:flint Docker image while the upstream merge remains pending (vLLM PR #62938)—and cites about 500 tokens/s for a single request on one GB200 even with 100K context—useful for teams that want the same weights locally and in a rack. For readers already shopping RTX Spark laptops for local models, Flint is an early software story that matches that hardware narrative.

Benchmarks (Recursive's numbers)

Recursive says Flint leads MoE models under 50B total parameters on agentic coding benchmarks—including SWE-Bench Pro v2, Terminal-Bench 2.1, and SWE-Bench Verified—and rivals dense 25–40B models while running more than five times faster (Recursive's comparison framing). The Hugging Face model card lists publisher-reported release evaluations:

• Terminal-Bench 2.1: 54.9%
• SWE-bench Pro v2: 89.7%
• SWE-bench Verified: 74.2%

Those figures are Recursive's own reporting, not independent verification in this article. They are still useful as a baseline for what the team claims at launch—and for how Flint positions against other open coding models you might run locally, such as large open-weight stacks discussed in our Muse Spark open-weights piece.

How Recursive built it with automated research

Recursive credits the Eureka Machine for architecture search, quantization (4-bit output layer, NVFP4 delivery), and kernels such as the Metal megakernel that fuses decode on M4 and M5 chips—optimizations aimed squarely at tokens per second.

How to run Flint locally

Download weights from Hugging Face. Supported stacks include llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang. For rec on Apple Silicon, you need M4 Pro, M5 Pro, or M5 Max, 48 GB+ unified memory, and about 30 GB free disk; Recursive points users to harnesses like pi and OpenCode. BF16 serving fits one B300 or H200 via SGLang; vLLM targets GB200 throughput per its Oct. 21 post. Treat Flint as a daily offline coding model, but split complex tasks and steer more than you would with larger closed agents.

Why local coding models matter

Agentic coding sends paths, diffs, and repos through inference—often sensitive data teams cannot route to a third-party API. Local models cut recurring API fees on long sessions and keep workloads on hardware you control, at the cost of RAM, disk, and ops. Flint does not replace sandboxing or review, but it is another open-weight option as local AI PCs and cloud agents compete for developers.

Bottom line

Flint pairs Apache 2.0 MoE weights, the rec engine, and day-0 vLLM support for developers who want fast, private coding agents on MacBooks and RTX Spark-class boxes. Outside replication of Recursive's benchmarks is still pending, but the launch is a concrete local option worth piloting on your own repos before you depend on it in production.

Sources

• Recursive: Introducing Flint — Oct. 21, 2026
• Hugging Face model card: RecursiveSI/flint
• GitHub: recursive-oss/rec
• vLLM Blog: Day-0 Support for Flint on vLLM — Oct. 21, 2026