39.8B total / approximately 2.5B active
262,144 tokens
Apache-2.0 for the model weights and Recursive's rec local inference engine; hosted API terms and data practices are separate
Apache-2.0 decoder-only Mixture-of-Experts language model designed for local coding agents that investigate repositories, fix bugs, implement features, write tests, reason, and use tools over multiple turns. Flint was designed and trained from scratch by Recursive Superintelligence in under four months with assistance from the company's Eureka Machine automated research system. The model contains 39.8B total parameters and activates approximately 2.5B parameters per token when counting the output head, or approximately 2.2B active body parameters. Its rounded designation is 40B-A2.5B. The 49-layer architecture begins with two dense layers followed by 47 MoE layers and combines 35 sliding-window attention layers with 14 full-attention layers. Sliding attention covers the current token plus 512 preceding tokens. Attention uses 32 query heads, four key/value heads, 128-dimensional heads, per-head QK RMS normalization, and learned attention sinks. Sliding-window layers use rotary position embeddings, while full-attention layers use no positional encoding. Each MoE layer contains 256 routed experts, selects six per token, and also activates one shared expert. The model has a 2,048-dimensional hidden state, 6,144-dimensional dense intermediate layers, a 1,024-dimensional shared expert, an untied 150,144-token vocabulary, and a stated June 2026 knowledge cutoff. The distributed configuration and serving recipes use a combined prompt-and-output context of 262,144 tokens. Recursive's launch post says the architecture supports up to 512K tokens but recommends staying within approximately 256K for coding quality; the canonical checkpoint configuration therefore uses 262,144. Flint exposes reasoning separately through reasoning_content, supports OpenAI-compatible tool calls, and recommends preserving reasoning when replaying conversation history. Publisher evaluations use temperature 1.0, top-p 0.95, top-k 20, and average pass@1 over at least five attempts. Recursive reports 54.9% on Terminal-Bench 2.1 using Terminus 2, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified using mini-swe-agent. Its Terminal-Bench evaluation treats solutions obtained through web retrieval as unsuccessful, and Recursive notes that comparison-model evaluation setups may differ. Training proceeded through pretraining, mid-training, supervised fine-tuning, and reinforcement learning. Disclosed training sources include publicly available and crawled web documents, text and code, commercially licensed data, AI-generated synthetic data, instruction tuning, reinforcement learning, and human-preference data. Recursive's Eureka Machine contributed to data construction, optimizer and training recipes, architecture search, infrastructure, quantization, and inference kernels. The BF16 checkpoint contains 39,785,287,456 parameters and approximately 79.58 GB of safetensors across 15 shards and is packaged with an SGLang serving kit for one H200 or B300 GPU per replica. The approximately 22.57 GB MLX INT4 checkpoint uses group-64 affine INT4 projections and experts with INT8 embeddings and targets Apple silicon through Recursive's rec engine. A 25.99 GB Q4_1 GGUF targets llama.cpp with Metal on Apple silicon, while a 22.86 GB NVFP4 GGUF targets NVIDIA GPUs and has been tested on B300, RTX PRO 6000 Blackwell, RTX Spark, and DGX Spark; both local profiles recommend at least 48 GB of memory for the documented full-context configuration. The quantization process includes the output head at four bits and uses offline distillation from the BF16 model over a self-generated coreset; Recursive reports that the final NVFP4 model remains within 0.7 benchmark points of BF16. The company reports more than 125 generated tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64 GB NVIDIA RTX Spark at short contexts using its optimized stack. Supported deployment paths include SGLang, vLLM, llama.cpp, Ollama, LM Studio, Unsloth, the lightweight rec inference engine, Pi, and OpenCode.