
Recursive dropped Flint last week: a 40-billion-parameter coding model that runs on a MacBook, benchmarks above most commercial alternatives on agentic tasks, and costs nothing per token. It’s Apache 2.0. It was largely built by Recursive’s own automated research system — the Eureka Machine — in under four months. That last part is either the most impressive milestone in AI this year or the most carefully worded press release, depending on how you read it. Either way, the model is real, the benchmarks are competitive, and you can run it today.
Running Flint in Under Five Minutes
If you have an Apple Silicon Mac with 48GB of unified memory or a GPU with 32GB+ VRAM, you’re ready to go. The official int4 quantization weighs 22.9GB — it clears most modern workstation GPUs with room to spare.
The simplest path is a single Ollama command:
ollama run hf.co/RecursiveSI/flint-int4-gguf:Q4_1
On Apple Silicon, Recursive ships its own inference engine — rec — with kernels optimized specifically for M4 and M5 chips. It outperforms Ollama on MacBooks:
rec launch pi
For llama.cpp users: you need at least nightly build b11803 (released October 21, 2026). Older builds will not load Flint correctly.
If 32GB is too tight, Unsloth’s 3-bit quantization gets the footprint to 18.5GB — that fits a 24GB RTX 4090-class GPU. Here’s the full quantization breakdown:
| Quantization | Size | Minimum RAM | Notes |
|---|---|---|---|
| BF16 (full) | 79.6 GB | 80 GB+ | Research / multi-GPU |
| Q6_K_L | 37 GB | 48 GB | Near-lossless quality |
| Q4_1 / NVFP4 | 23–26 GB | 32 GB | Best practical choice |
| 3-bit (Unsloth) | 18.5 GB | 24 GB | Fits RTX 4090 |
The Benchmarks: Honest Assessment
Recursive reports 74.2% on SWE-bench Verified and 89.7% on SWE-bench Pro v2. The SWE-bench Verified score is the one to trust — it’s the more audited benchmark, and 74.2% is legitimately strong for a sub-50B model that runs on your own hardware. Flint beats every other model in that parameter class on that measure.
For context: DeepSeek V4-Pro hits 80.6% on SWE-bench Verified, and Qwen3.8-27B scores higher on Recursive’s own six-test aggregate. Both of those require cloud access. Flint’s SWE-bench Pro v2 score of 89.7% uses Recursive’s own benchmark — treat it as directionally meaningful until third-party audits arrive.
Speed is where the story gets interesting. On an M4 Pro MacBook, Flint generates 131 tokens per second. That’s roughly double comparable local models. The reason comes down to architecture.
Why a 40B Model Runs This Fast
Flint is a mixture-of-experts (MoE) model. It has 39.8 billion total parameters distributed across 256 “experts,” but only 2.5 billion are active for any given token. The routing layer selects 6 experts per token and ignores the rest. Your GPU processes 2.5B active parameters per step — not 40B. That’s why the speed is real.
The architecture is 49 layers: two dense layers, then 47 MoE layers. Most layers use sliding-window attention for efficiency; a subset uses full attention for longer-range reasoning. Context window: up to 512k tokens, with 256k recommended for optimal performance.
The Eureka Machine: AI Building AI
Recursive’s Eureka Machine — an automated research loop — reportedly handled architecture search, training recipe design, quantization strategy, and custom inference kernel development for Flint. The TechCrunch launch piece notes the company says it “contributed to nearly every part of building Flint.” CEO Richard Socher, former Salesforce chief scientist, frames the long-term goal as “full recursive self-improvement, with humans overseeing the process.”
That framing deserves scrutiny. “Contributed to” and “built by” are different claims. The Eureka Machine proposes experiments, runs them, and surfaces results — but humans evaluate outputs, make architectural decisions, and integrate findings. This is AI-accelerated research, not autonomous AI self-improvement. Still impressive. Not science fiction yet.
When to Use Flint — and When to Stay on Copilot
Flint earns its place when privacy is the constraint. Regulated industries — fintech, healthcare, defense — where source code cannot touch third-party servers now have a real option. It also makes sense when API costs have become a budget line worth cutting. GitHub Copilot runs about $19/month per developer; Cursor Pro is $20/month. A 20-person team spends roughly $4,800/year. If the hardware already exists, Flint’s cost is zero.
It is not a drop-in replacement for IDE-integrated agents today. Flint performs best with explicit, focused instructions — “fix the authentication bug in auth.py lines 45–67″ beats “fix the security issues in this codebase.” That’s a workflow adjustment, not a dealbreaker, but worth knowing before you cancel your subscription.
The verdict: Flint is the best free, local, Apache 2.0 coding agent available in October 2026. The install is one command if you have the hardware. The benchmarks are strong, the speed is real, and the cost is zero. Check the GGUF releases on Hugging Face for the quantization that matches your setup, and run it before your next Copilot renewal.

