Recursive Superintelligence has released Flint, its first model. It is a mixture-of-experts (MoE) architecture with 40 billion total parameters but only 2.5 billion active. The team states that an automated research system called the Eureka Machine built nearly every part of the stack. The weights are available on Hugging Face under an Apache 2.0 licence.
Specs and performance
Flint has 39.8 billion total parameters and runs with about 2.5 billion active per token. The model card documents a context window of 262,144 tokens, though the blog post cites up to 512k.
You can run it on Apple Silicon Macs with 48 GB RAM, specifically the M4 Pro, M5 Pro, or M5 Max. It also works on 48 GB NVIDIA GPUs. 4-bit quantised weights take approximately 23 GB of memory.
Among MoE models under 50 billion parameters, it tops Recursive’s chart for agentic coding. It trails the strongest dense models. Its best scores are 95.1% on IFEval and 74.2% on SWE-bench Verified. Its worst result is 54.9% on Terminal-Bench 2.1, compared to 73.0% for dense Qwen 3.8.
What is Recursive Flint?
Flint is a decoder-only MoE Transformer designed for coding agents. It investigates codebases, fixes bugs, implements features, and writes tests. The model is text-only and supports reasoning alongside OpenAI-compatible tool calls.
Introducing Flint, @Recursive_SI‘s first model: a fast local coding model built with our automated research system, the Eureka Machine.
• Best-in-class fast local coding
• Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark
• Built from… pic.twitter.com/ufg2xScSGw— Richard Socher (@RichardSocher) October 21, 2026
How does Flint’s architecture work?
The model has 49 layers in total: two dense layers followed by 47 MoE layers. Each MoE layer contains 256 routed experts. A router selects six experts per token, plus one always-active shared expert.
Attention mixes two types. Thirty-five layers use a 513-token sliding window. Fourteen layers use full attention with no positional encoding. There are 32 query heads and four key/value heads.
The 2.5 billion active figure includes the output head. The active body alone is about 2.2 billion 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 three contributions:
- Architecture: It started from a 35 billion total, 3 billion active baseline. It searched expert configurations and widths, landing on 40 billion total and 2.5 billion active.
- Quantisation: It quantised 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 quantisation-aware training baseline. The NVFP4 model scores within 0.7 points of the 16-bit model.
- Kernels: It wrote specialised 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, Qwen 3.6 scores 73.4% on SWE-bench Verified. Dense Qwen 3.8 scores 73.0% on Terminal-Bench 2.1, well above Flint.
Comparison with Qwen 3.6, Laguna XS 2.1 and Gemma 4
Competitor figures below come from each developer’s own model card, not Recursive’s chart.
| Feature | Recursive Flint | Qwen 3.6-35B-A3B | Laguna XS 2.1 | Gemma 4 31B |
|---|---|---|---|---|
| Developer | Recursive | Qwen (Alibaba) | Poolside | Google DeepMind |
| Total params | 39.8B | 35B | 33B | 30.7B |
| Active params | ~2.5B | 3B | 3B | 30.7B (dense) |
| Context | 262,144 (blog: up to 512k) | 262,144 native, up to 1,010,000 | 262,144 | 256K |
| Modality | Text | Text + image | Text | Text + image |
| License | Apache 2.0 | Apache 2.0 | OpenMDW-1.1 | Apache 2.0 |
| Local hardware | 48 GB Mac or GPU | Not disclosed | Mac with 36 GB RAM | Consumer GPUs, workstations |
| SWE-bench Verified | 74.2% | 73.4% | 70.9% | Not disclosed |
| Terminal-Bench | 54.9% (v2.1) | 51.5% (v2.0) | 37.5% (v2.0) | Not disclosed |
| LiveCodeBench v6 | 80.3% | 80.4% | Not disclosed | 80.0% |
| SWE-bench Pro | 89.7% (v2) | 49.5% (public) | 47.6% (public) | Not disclosed |
Hardware and speed
Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro. It reports more than 150 tokens per second on a 64 GB NVIDIA RTX Spark. Both are short-context figures, up to 2x faster than comparable models. It also claims Flint runs more than 5x faster than dense 25-40B models with similar coding scores.
There are three local options:
- Apple Silicon: Flint INT4 MLX weights take about 22.6 GB. The rec engine needs an M4 Pro, M5 Pro or M5 Max with 48 GB memory. It reports 1.4 to 2.6x the throughput of llama.cpp on an M4 Pro.
- NVIDIA: The NVFP4 GGUF file is 22.9 GB. Recursive recommends at least 48 GB of GPU memory.
- Server: The 80 GB BF16 checkpoint needs one B300 or H200 GPU per replica.