Recursive Releases Open-Weight Coding Model Flint, a 40B MoE with Only 2.5B Active Parameters
Recursive Superintelligence has released its first model, Flint, a 40B MoE model aimed at coding agents with roughly 2.5B active parameters per token; the weights are released under the Apache 2.0 license on Hugging Face. The model card lists a context of 262,144 tokens, with 4-bit weights at about 23 GB, runnable on a 48 GB Apple Silicon Mac or a 48 GB NVIDIA GPU. The company self-reports SWE-bench Verified at 74.2% and IFEval at 95.1%, while Terminal-Bench 2.1 stands at 54.9%, below dense Qwen 3.8's 73.0%. Recursive says its automated research system Eureka Machine contributed to the design of the architecture, quantization, and kernels.
Why read this
AI score 70
The original article provides the local deployment configuration for the 40B MoE coding model on 48GB devices along with comparisons across multiple benchmarks, which can be used to gauge the hardware threshold for local coding agents.
Topics
Body
Add as a preferredsource on Google
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.
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?
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.
BenchmarkFlint BF16Flint NVFP4
Terminal-Bench 2.154.9%53.7%
SWE-bench Pro v289.7%88.5%
SWE-bench Verified74.2%74.1%
LiveCodeBench v680.3%80.8%
IFBench89.2%88.1%
IFEval95.1%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):
ModelTerminal-Bench 2.1SWE-bench Pro v2†SWE-bench VerifiedLiveCodeBench v6IFBenchIFEval
Flint BF1654.989.774.280.389.295.1
Flint NVFP453.788.574.180.888.194.6
Qwen 3.5 (MoE)38.768.669.274.670.291.9
Qwen 3.6 (MoE)44.973.873.480.467.989.6
Laguna XS 2.1 (MoE)24.7†38.670.949.5†34.4†73.7†
Qwen 3.8 (dense)73.094.975.6†90.379.594.6†
Muse Glimmer (dense)51.786.076.063.5†77.094.7†
Gemma 4 (dense)43.461.766.680.076.094.6
How does Flint compare with Qwen3.6, Laguna XS 2.1 and Gemma 4?
Competitor figures below come from each developer’s own model card, not Recursive’s chart.
FeatureRecursive FlintQwen3.6-35B-A3BLaguna XS 2.1Gemma 4 31B
DeveloperRecursiveQwen (Alibaba)PoolsideGoogle DeepMind
Total params39.8B35B33B30.7B
Active params~2.5B3B3B30.7B (dense)
Context262,144 (blog: up to 512k)262,144 native, up to 1,010,000262,144256K
ModalityTextText + imageTextText + image
LicenseApache 2.0Apache 2.0OpenMDW-1.1Apache 2.0
Local hardware48 GB Mac or GPUNot disclosedMac with 36 GB RAMConsumer GPUs, workstations
SWE-bench Verified74.2%73.4%70.9%Not disclosed
Terminal-Bench54.9% (v2.1)51.5% (v2.0)37.5% (v2.0)Not disclosed
LiveCodeBench v680.3%80.4%Not disclosed80.0%
SWE-bench Pro89.7% (v2)49.5% (public)47.6% (public)Not disclosed
How fast is Flint, and what hardware does it need?
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 3 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 1 B300 or H200 GPU per replica.
Flint also works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang.
Key Takeaways
Flint is a 40B MoE coding model with only 2.5B active parameters.
It scores 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1.
4-bit weights run in about 23 GB on a 48 GB Mac or GPU.
Recursive’s Eureka Machine helped design the architecture, quantization and kernels.
SWE-bench Pro v2 scores are Recursive-measured and not comparable to public SWE-Bench Pro.
Asif RazzaqWebsite | + postsBio ⮌
Asif RazzaqGoogle Cloud Ships llm-d 1.0 With Cache-Aware Routing for TPU and GPU Fleets
Asif RazzaqDeepSeek Open-Sources DeepInfer, the Inference Framework Behind DeepSeek-V4.1-Flash
RELATED ARTICLESMORE FROM AUTHOR
Google Cloud Ships llm-d 1.0 With Cache-Aware Routing for TPU and GPU Fleets
DeepSeek Open-Sources DeepInfer, the Inference Framework Behind DeepSeek-V4.1-Flash
Exa 3.0 Search API Tops Perplexity, Brave and Parallel on 8 Benchmarks