Recursive Flint Review: Local Coding Model Features, Benchmarks and Deployment Options

Recursive Flint delivers an open-weight decoder-only Mixture-of-Experts model built for local coding workflows. Released on October 21, 2026 by Recursive Superintelligence, the model targets codebase investigation, bug fixing, feature implementation and test generation without cloud dependencies.

The company, which emerged from stealth in May 2026 with $650 million in funding, positions Recursive Flint as a test of recursive self-improvement. Its knowledge cutoff sits at June 2026 and the model ships under an Apache 2.0 license.

Recursive Flint Model Architecture and Capabilities

The model contains 39.8 billion total parameters with roughly 2.5 billion active per token. It uses 256 routed experts, selecting six plus one shared expert per MoE layer across 49 total layers. Hybrid attention combines 35 sliding-window layers of 513 tokens with 14 full-attention layers, supporting a documented context of 262,144 tokens and up to 512k in testing.

Input and output remain text-only, with built-in reasoning and multi-turn tool use that follows OpenAI-compatible formats. Users can run the model for terminal commands, code editing and agentic harnesses such as Pi, OpenHands or VS Code extensions. Quantized 4-bit versions occupy about 23 GB, fitting on 48 GB unified-memory Macs or comparable NVIDIA GPUs.

Inference speeds exceed 125 tokens per second on an M4 Pro MacBook Pro and 150 tokens per second on 64 GB NVIDIA RTX hardware for short contexts. The weights appear on Hugging Face under RecursiveSI/flint alongside BF16 checkpoints and optimized GGUF, MLX and INT4 files.

Benchmarks, Deployment and Limitations

Vendor-reported scores include 54.9 percent on Terminal-Bench 2.1, 89.7 percent on SWE-bench Pro v2 and 74.2 percent on SWE-bench Verified. Additional results show 80.3 percent on LiveCodeBench v6, 89.2 percent on IFBench and 95.1 percent on IFEval. These figures come from Recursive’s own evaluations and remain unverified by independent leaderboards as of October 23, 2026.

Deployment requires no per-token fees once the model runs locally. Supported runtimes include Recursive’s rec engine for Macs, llama.cpp, Ollama, LM Studio, Unsloth, vLLM and SGLang. Hosted rate-limited access exists on OpenRouter and OpenCode, but the emphasis stays on offline private use. The 4-bit checkpoint plus supporting files total roughly 22–26 GB.

Early feedback notes that the model can require step-by-step task breakdown compared with larger closed models. No safety evaluations appear on the model card, and all performance numbers originate from the developer. Users should review generated code before execution, particularly for larger projects.

Compared with Qwen 3.8 27B, Recursive Flint trails on several dense-model benchmarks yet leads among other MoE models under 50 billion parameters on the reported agentic coding suites. The absence of third-party verification and limited published safety data remain the primary constraints at launch.