Recursive Releases Flint, a 40B MoE Coding Model That Runs Locally at 125 Tokens per Second

Recursive Superintelligence has released Flint, its first model: a mixture-of-experts (MoE) coding model with 39.8B total parameters and roughly 2.5B active per token, published on Hugging Face under an Apache 2.0 license. The company states that its Eureka Machine, an automated research system that proposes ideas, implements them, tests them and picks what to try next, helped build nearly every part of the stack — architecture, quantization and inference kernels.

The release matters for two reasons. First, it pushes competitive agentic coding (models that investigate codebases, fix bugs, implement features and write tests autonomously) onto consumer hardware: Flint runs on a 48 GB Apple Silicon Mac or a 48 GB NVIDIA GPU. Second, it is an early, public test of whether automated AI research systems can design competitive models with minimal human intervention.

What can Flint do?

Flint is a decoder-only MoE Transformer built for coding agents. It is text-only, supports reasoning and OpenAI-compatible tool calls. Its documented context window is 262,144 tokens on the model card; the announcement blog cites up to 512k.

Recursive's headline claims, posted by founder Richard Socher on October 21, 2026, center on speed: "Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark." The company also claims Flint runs more than 5x faster than dense 25–40B models with similar coding scores.

Key specifications:

• Total parameters: 39.8B; ~2.5B active per token, including the output head (the body alone is about 2.2B)
• Context: 262,144 tokens documented, up to 512k per the blog
• Quantized footprint: about 23 GB for 4-bit weights
• License: Apache 2.0

How does Flint's architecture work?

The model has 49 layers: 2 dense layers followed by 47 MoE layers. Each MoE layer contains 256 routed experts. A router selects 6 experts per token, plus 1 always-active shared expert.

Attention mixes two types across the network. 35 layers use a 513-token sliding window (attention restricted to a fixed local span), while 14 layers use full attention with no positional encoding. The attention configuration includes 32 query heads and 4 key/value heads.

How did the Eureka Machine build Flint?

Recursive attributes three concrete contributions to its automated research system.

Architecture search. The Eureka Machine started from a 35B-total, 3B-active baseline. It searched expert configurations and widths, landing on the final 40B total and 2.5B active configuration.

Quantization. It quantized the output layer (lm_head, the component that maps hidden states to vocabulary tokens) to 4 bits — against common practice. An offline distillation stage, using the 16-bit model as teacher, recovered accuracy. Recursive says this approach nearly halved training steps compared with a quantization-aware training baseline. The resulting NVFP4 model scores within 0.7 points of the 16-bit model.

Kernels. The system wrote specialized kernels for low-bit projections, MoE operations, attention and prompt processing. On Apple M4 and M5 chips, a Metal megakernel fuses the entire decode forward pass into one operation.

How does Flint perform on coding benchmarks?

All scores below are self-reported by Recursive, sampled at temperature 1.0, top-p 0.95 and top-k 20. Agentic scores average pass@1 over at least 5 attempts.

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%

Benchmark methodology carries caveats. Terminal-Bench used the Terminus 2 harness, and solutions found through web retrieval counted as failures. SWE-bench runs used mini-swe-agent with no internet access. Recursive flags that its SWE-bench Pro v2 scores are internally measured and not comparable to public SWE-Bench Pro results.

The strongest scores are IFEval at 95.1% (instruction following) and SWE-bench Verified at 74.2% (real-world bug fixing). The weakest is Terminal-Bench 2.1 at 54.9%, where dense Qwen 3.8 scores 73.0% — a gap of more than 18 points.

How does Flint compare with Qwen3.6, Laguna XS 2.1 and Gemma 4?

Within Recursive's comparison chart, Flint tops every MoE model under 50B on agentic coding, but trails the strongest dense models. Qwen3.6 (MoE) scores 73.4% on SWE-bench Verified, within 0.8 points of Flint's 74.2%. Qwen 3.8 (dense) leads broadly: 73.0% on Terminal-Bench 2.1, 94.9% on SWE-bench Pro v2 and 90.3% on LiveCodeBench v6.

Against developer-reported model-card figures:

Feature	Recursive Flint	Qwen3.6-35B-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

On Terminal-Bench, Recursive reports 54.9% on v2.1, versus Qwen3.6's 51.5% and Laguna XS 2.1's 37.5% — both on v2.0, which limits direct comparison. On LiveCodeBench v6, Flint's 80.3% sits between Qwen3.6's 80.4% and Gemma 4's 80.0%.

How fast is Flint, and what hardware does it need?

Recursive reports more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64 GB NVIDIA RTX Spark. Both figures are short-context measurements.

Three local deployment options exist:

• Apple Silicon: Flint INT4 MLX weights take about 22.6 GB. The open-source rec engine requires an M4 Pro, M5 Pro or M5 Max with 48 GB of memory, and 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.

Flint also runs through llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang, which extends its reach beyond Recursive's own tooling.

What is the so-what?

Flint establishes two things at once: that a 2.5B-active MoE can reach 74.2% on SWE-bench Verified while running entirely on a laptop, and that an automated research system can plausibly drive architecture search, quantization strategy and kernel engineering. The open question is whether the Eureka Machine can repeat the feat at larger scale — and whether self-reported benchmark wins, particularly on internally measured suites like SWE-bench Pro v2, hold up under independent evaluation.