Memory at each precision
| Version | Weights | With context | Smallest Mac |
|---|---|---|---|
| 8-bit | 42.3 GB | 45.7 GB | 64 GB |
| 6-bit | 32.3 GB | 34.9 GB | 48 GB |
| 4-bit | 22.4 GB | 24.2 GB | 36 GB |
| 3-bit lower quality | 17.4 GB | 18.9 GB | 32 GB |
| 2-bit lower quality | 12.4 GB | 13.9 GB | 24 GB |
On every Apple chip
The smallest memory size of each chip that runs Recursive Flint, the best version that fits it, and the writing speed you can expect.
| Chip | Needs | Version | Tokens/s |
|---|---|---|---|
| M6 | 24 GB | 2-bit Runs well | 96–170 |
| M5 Ultra | 96 GB | 8-bit Runs well | 200–356 |
| M5 Max | 36 GB | 4-bit Runs well | 148–263 |
| M5 Pro | 48 GB | 6-bit Runs well | 69–123 |
| M5 | 24 GB | 2-bit Runs well | 86–153 |
| M4 Max | 36 GB | 4-bit Runs well | 132–234 |
| M4 Pro | 48 GB | 6-bit Runs well | 61–109 |
| M4 | 24 GB | 2-bit Runs well | 68–120 |
| M3 Ultra | 96 GB | 8-bit Runs well | 137–243 |
| M3 Max | 36 GB | 4-bit Runs well | 96–171 |
| M3 Pro | 36 GB | 4-bit Runs well | 48–86 |
| M3 | 24 GB | 2-bit Runs well | 56–100 |
| M2 Ultra | 64 GB | 8-bit Runs well | 133–237 |
| M2 Max | 64 GB | 8-bit Runs well | 67–119 |
| M2 Pro | 32 GB | 3-bit Runs well | 82–145 |
| M2 | 24 GB | 2-bit Runs well | 56–100 |
| M1 Ultra | 64 GB | 8-bit Runs well | 133–237 |
| M1 Max | 64 GB | 8-bit Runs well | 67–119 |
| M1 Pro | 32 GB | 3-bit Runs well | 82–145 |
| M1 | 16 GB | 2-bit Only just | 38–68 |
Measured by others
Decode speeds for Recursive Flint that other people measured on their own Macs and posted publicly. These are third-party measurements, not ours, and not the estimates above.
| Mac | Runtime | Precision | Decode tok/s | Source |
|---|---|---|---|---|
| Mac Studio M3 Ultra | Ollama (MLX engine) | INT4 (QAT) | 107median of 7 prompts × 3 runs, reasoning on | Alec Kloss @aleckloss |
| Mac Studio M3 Ultra | mlx-lm (PR #1998) | INT4 (QAT) | 104.5median of 5 runs, 19-token prompt, 512 tokens out | Alec Kloss @aleckloss |
| Mac Studio M3 Ultra | mlx-vlm | BF16 | 51.3at 1K context | Prince Canuma @Prince_Canuma |
| Mac Studio M3 Ultra | mlx-vlm | BF16 | 33at 128K context | Prince Canuma @Prince_Canuma |
| Mac mini M5 Pro 64 GB | Ollama | 4-bit | 115short prompt | Pratik Gandhi @pratikg |
| Mac mini M5 Pro 64 GB | Ollama | 4-bit | 87at 20K contextafter reading a 20K-token codebase (prefill took 11 s) | Pratik Gandhi @pratikg |
Results vary by runtime and settings: its version, the precision, the context length, the prompt and reasoning on or off all change the speed.
How to run it
- In Terminal with Rec, the maker’s own engine:
curl -fsSL https://raw.githubusercontent.com/recursive-oss/rec/main/install.sh | bashrec launch piNeeds: M4 Pro, M5 Pro or M5 Max · at least 48 GB unified memory · about 30 GB free disk · Node.js 22 and npm (for pi)Build: RecursiveSI/flint-int4 on Hugging Face. Other precisions sit next to it.
Step by step: download and run LLMs on a Mac · how to use MLX Harness · MLX vs Ollama vs llama.cpp
About Recursive Flint
| Made by | Recursive |
|---|---|
| Parameters | 39.8B total · 2.5B active (mixture of experts) |
| Context | 256K tokens |
| Licence | Apache-2.0 |
| Released | 2026-10 |
| Input | Text |
| Weights | RecursiveSI/flint |
Questions
- How much memory does Recursive Flint need?
- About 22.4 GB of weights at 4-bit, plus room for the context: 24.2 GB in all. On a Mac that means 36 GB of unified memory or more, since macOS lets the GPU use about 75% of it by default. Recursive’s own engine, Rec, needs an M4 Pro, M5 Pro or M5 Max with at least 48 GB of unified memory (maker’s requirement).
- How fast is Recursive Flint on a Mac?
- Writing speed depends on memory bandwidth. The fastest Mac for it here is an M5 Ultra with 96 GB at about 200–356 tokens a second (8-bit). The smallest Mac that runs it, with 36 GB, manages about 148–263 on an M5 Max.
- Why does Recursive Flint run faster than its size suggests?
- It is a mixture-of-experts model: it holds 39.8B parameters in memory but reads only 2.5B for each token, so it writes at the speed of a much smaller model.
Similar models
For your model card or README
Shows the smallest Mac that runs Recursive Flint well and links here.
[](https://mlx.app/run/flint)How it’s worked out
Memory: the model’s parameters at 8, 6, 4, 3 or 2 bits per weight plus MLX’s group scales (half a bit per weight), or its native format, with room for the context on top. By default macOS lets the GPU use about 75% of unified memory. Speed: memory bandwidth divided by the bytes read per token is the ceiling for one conversation; in llama.cpp’s public Apple Silicon benchmarks real runs reach about 45% to 80% of it, which is the range shown. These are estimates, not measurements: measure your own Mac on MLX Bench. Model figures from each official model card, chip figures from Apple’s Tech Specs. Checked 2026-10-19. Same data as mlx.app/compare.
Open data: mlx.app/data (JSON, CSV)