flint-MXFP4-MoE-40B-A2.5B-GGUF

Recursive Flint (40B-A2.5B) is an open-weight, Apache 2.0-licensed decoder-only mixture-of-experts (MoE) language model developed from scratch by Recursive (Recursive Superintelligence, Inc.) specifically for coding agents. Built with 39.8 billion total parameters and approximately 2.5 billion active parameters per token across 49 layers (2 dense and 47 MoE layers using a mix of 513-token sliding-window and full attention), the model supports a 262,144-token context window, extended reasoning, and multi-turn OpenAI-compatible tool use. Purpose-built to investigate codebases, fix bugs, implement features, and write tests, Flint achieves 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified, and 54.9% on Terminal-Bench 2.1. While the main repository provides the BF16 checkpoint (~80 GB) and an SGLang serving kit tailored for single NVIDIA B300 or H200 GPUs, separate quantized variants (such as NVFP4, MLX INT4, and GGUF) are also available for local execution on NVIDIA hardware and Apple Silicon.

Model Files

File Name	Quant Type	File Size	File Link	Description
flint.BF16.gguf	BF16	79.6 GB	Link	Full BF16 weights. Highest quality, largest file size.
flint.MXFP4_MOE.gguf	MXFP4_MOE	22.2 GB	Link	Mixed precision / MoE optimized quantization. Highly efficient size for massive models, recommended if your inference engine supports it.
flint.Q5_K_M.gguf	Q5_K_M	28.3 GB	Link	High quality, recommended.

MXFP4-MOE Test Run

Tested flint.MXFP4_MOE on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU with 96 GB of VRAM, using llama-cli with full GPU offloading.

Performance Metrics

Metric	Value
Prompt Processing Speed	262.8 tokens/s
Generation Speed	164.9 tokens/s
Total Wall Time	14.231 seconds
Context Length	8,192 tokens
KV Cache (K)	`q8_0`
GPU	NVIDIA RTX PRO 6000 Blackwell Server Edition
GPU VRAM Capacity	96 GB
Peak GPU VRAM Usage	21.35 GiB
Peak GPU Utilization	92%
GPU Layers Offloaded	99 (`-ngl 99`)
Maximum Generation Tokens	16,384
Random Seed	42

Run Inference

./build/bin/llama-cli \
  -m /path/to/models/flint.MXFP4_MOE.gguf \
  -ngl 99 \
  -c 8192 \
  --cache-type-k q8_0 \
  --seed 42 \
  -n 16384 \
  -st \
  -p "Your prompt here"

Notes

Full GPU offloading: -ngl 99 requests offloading up to 99 model layers to the GPU. Actual offloading depends on the model architecture and the number of layers available.
KV cache: --cache-type-k q8_0 configures the K cache to use the q8_0 quantization format.
Context length: -c 8192 sets the context size to 8,192 tokens.
Generation length: -n 16384 sets the maximum number of tokens to generate.
Random seed: --seed 42 sets the random seed for reproducibility.

Performance results reflect this specific hardware and inference configuration. Actual throughput may vary depending on the prompt, model, software build, and runtime settings.

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

Quantized using llama.cpp by prithivMLmods