Recursive releases Flint, an Apache 2.0 open-weight 40B-A2.5B MoE for local coding agents: self-reported 74.2% SWE-bench Verified and 150+ tokens/s on a 64GB RTX Spark, built with its Eureka Machine automated-research system

Why it matters

【Key points】

Richard Socher's Recursive released its first model, Flint, on Oct 21: a from-scratch MoE for local coding agents, Apache 2.0 weights on Hugging Face.

1. Specs: 39.8B total, ~2.5B active; 35 of 49 layers use a 513-token sliding window; 256 experts (top-6 + shared); 262K context.

2. Self-reported scores: Terminal-Bench 2.1 54.9%, SWE-bench Pro v2 89.7%, SWE-bench Verified 74.2%. Claimed best among sub-50B MoEs.

3. Speed: 125+ tok/s on an M4 Pro MacBook Pro, 150+ on a 64GB RTX Spark; vLLM day-0 ~485 tok/s on one GB200 at 100K context.

4. Built in under four months with its Eureka Machine automated-research system shaping architecture, quantization (NVFP4 within 0.7 points of BF16) and kernels.

5. Runs on llama.cpp, Ollama, LM Studio, MLX, vLLM, SGLang; NVIDIA listed it as an RTX Spark launch model.

【Verifiable】

https://www.recursive.com/updates/introducing-recursive-flint

https://huggingface.co/RecursiveSI/flint

https://vllm.ai/blog/2026-10-21-flint

https://x.com/RichardSocher/status/2112944391266386133

【Our read】

1. Background: Recursive emerged in May with $650M; Flint is its first testable product and a rare concrete case of AI-assisted model building.

2. Comparison: peers for local agents include Liquid's LFM2.5-24B-A2B, NVIDIA Nemotron 3.5 Bolt 74B-A9B, Meta's Muse Glimmer 30B and Perplexity's local PPLX model. Flint activates fewer parameters yet claims higher coding scores.

3. Impact: if tests hold, a 64GB laptop runs a capable coding agent offline, boosting local AI hardware and pressuring small Qwen/GLM coders.

4. Caveats: scores are self-reported; 89.7% on SWE-bench Pro v2 is unusually high. Await independent tests.