← CanIRunLocal · All hardware · All models

Can I run Flint 40B-A2.5B (MoE) locally?

Recursive · 39.8B parameters (2.5B active) · coding.

Ollama weight files: NVFP4 23 GB (flint:40b); Q4_1 26 GB (flint:40b-a2b-q4_1); INT4 23 GB (flint:40b-a2b-qat-int4).

flint:latest and flint:40b-a2b are the same NVFP4 file as flint:40b (23GB on the Ollama tag list). flint:40b-a2b-q4_1 is the 26GB Q4_1 GGUF and flint:40b-a2b-qat-int4 is the 23GB QAT INT4. Ollama also publishes Apple MLX tags flint:40b-a2b-gguf-nvfp4 (23GB) and flint:40b-a2b-mlx-bf16 (80GB); those MLX twins are left out of the fit list, same as other models. The Hugging Face BF16 checkpoint (RecursiveSI/flint) is about 80GB. Released 2026-10-21, Apache-2.0.

Yes on 47 machines in this list. The smallest is RTX 4060 Ti 16GB: NVFP4, about 24.6 GB, roughly 27 tokens per second, via Ollama. ollama run flint:40b

• DGX Spark 128GB: Q4_1, ~27.6 GB, fits in fast memory, ~96 tok/s — ollama run flint:40b-a2b-q4_1
• Mac Studio M5 Max 128GB: Q4_1, ~27.6 GB, fits in fast memory, ~150 tok/s — ollama run flint:40b-a2b-q4_1
• Mac Studio M5 Ultra 256GB: Q4_1, ~27.6 GB, fits in fast memory, ~294 tok/s — ollama run flint:40b-a2b-q4_1

Setup on DGX Spark 128GB (NVIDIA, ASUS, Dell, Lenovo, HP, MSI, Acer, Gigabyte)

• Ollama: ollama run flint:40b-a2b-q4_1

Hardware that fits

Estimates, not benchmarks. Model sizes are ollama.com download sizes; memory needed adds headroom for ~8k tokens of context; speeds are rough figures from memory bandwidth. Data updated 2026-10-22.

A project by OnDevice Labs · llms.txt