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. The weights are on Hugging Face under Apache 2.0.
TL;DR
- Size: 39.8B total, about 2.5B active per token. Context: 262,144 tokens documented on the model card; the blog cites up to 512k.
- Runs on: 48 GB Apple Silicon Macs (M4 Pro, M5 Pro, M5 Max) or 48 GB NVIDIA GPUs. 4-bit weights take about 23 GB.
- Performance: Tops Recursive’s chart among MoE models under 50B on agentic coding. Trails the strongest dense models.
- Best: IFEval 95.1%. SWE-bench Verified 74.2%.
- Worst: Terminal-Bench 2.1 at 54.9%, versus 73.0% for dense Qwen 3.8.
What is Recursive Flint?
Flint is a decoder-only MoE Transformer for coding agents. The model investigates codebases, fixes bugs, implements features and writes tests. It is text-only and supports reasoning and OpenAI-compatible tool calls.
How does Flint’s architecture work?
Flint has 49 layers: 2 dense layers, then 47 MoE layers. Each MoE layer has 256 routed experts. A router selects 6 per token, plus 1 always-active shared expert.
Attention mixes 2 types. 35 layers use a 513-token sliding window. 14 layers use full attention with no positional encoding. There are 32 query heads and 4 key/value heads.
The 2.5B active figure includes the output head. The active body alone is about 2.2B parameters.
How did the Eureka Machine build Flint?
The Eureka Machine proposes ideas, implements them, tests them and picks what to try next.
Recursive highlights 3 contributions:
- Architecture: It started from a 35B-total, 3B-active baseline. It searched expert configurations and widths, landing on 40B total and 2.5B active.
- Quantization: It quantized the output layer (lm_head) to 4 bits, against common practice. An offline distillation stage, with the 16-bit model as teacher, recovered accuracy. This nearly halved training steps versus a quantization-aware training baseline. The NVFP4 model scores within 0.7 points of the 16-bit model.
- Kernels: It wrote specialized kernels for low-bit projections, MoE operations, attention and prompt processing. On M4 and M5 chips, a Metal megakernel fuses the entire decode forward pass.
How does Flint perform on coding benchmarks?
All scores below are self-reported by Recursive. Sampling used temperature 1.0, top-p 0.95 and top-k 20.
| Benchmark | Flint BF16 | Flint NVFP4 |
|---|---|---|
| Terminal-Bench 2.1 | 54.9% | 53.7% |
| SWE-bench Pro v2 | 89.7% | 88.5% |
| SWE-bench Verified | 74.2% | 74.1% |
| LiveCodeBench v6 | 80.3% | 80.8% |
| IFBench | 89.2% | 88.1% |
| IFEval | 95.1% | 94.6% |
Terminal-Bench used the Terminus 2 harness. Solutions found through web retrieval counted as failures. SWE-bench runs used mini-swe-agent with no internet access. Agentic scores average pass@1 over at least 5 attempts.
In Recursive’s chart, Qwen3.6 scores 73.4% on SWE-bench Verified. Dense Qwen 3.8 scores 73.0% on Terminal-Bench 2.1, well above Flint.
Recursive’s own comparison chart (scores in %, † measured by Recursive, others developer-reported):
| Model | Terminal-Bench 2.1 | SWE-bench Pro v2† | SWE-bench Verified | LiveCodeBench v6 | IFBench | IFEval |
|---|---|---|---|---|---|---|
| Flint BF16 | 54.9 | 89.7 | 74.2 | 80.3 | 89.2 | 95.1 |
| Flint NVFP4 | 53.7 | 88.5 | 74.1 | 80.8 | 88.1 | 94.6 |
| Qwen 3.5 (MoE) | 38.7 | 68.6 | 69.2 | 74.6 | 70.2 | 91.9 |
| Qwen 3.6 (MoE) | 44.9 | 73.8 | 73.4 | 80.4 | 67.9 | 89.6 |
| Laguna XS 2.1 (MoE) | 24.7† | 38.6 | 70.9 | 49.5† | 34.4† | 73.7† |
| Qwen 3.8 (dense) | 73.0 | 94.9 | 75.6† | 90.3 | 79.5 | 94.6† |
| Muse Glimmer (dense) | 51.7 | 86.0 | 76.0 | 63.5† | 77.0 | 94.7† |
| Gemma 4 (dense) | 43.4 | 61.7 | 66.6 | 80.0 | 76.0 | 94.6 |
