Skip to content

Flint VRAM requirements

How much VRAM Flint needs
4-bit weights: 22.5 GBWith 8K context: 23.5 GBWith 32K context: 24.2 GB8-bit, 8K context: 40.4 GB
Parameters
39.8B
Active per token
1.7B
Longest context
256K
Released
Oct 2026
License
Apache 2.0
Strongest at
Not rated

How much VRAM Flint needs

Memory Flint needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit74.1 GB75.2 GB75.8 GB78.4 GB
8-bit39.4 GB40.4 GB41.1 GB43.7 GB
6-bit30.4 GB31.4 GB32.1 GB34.7 GB
5-bit26.4 GB27.4 GB28.1 GB30.7 GB
4-bit22.5 GB23.5 GB24.2 GB26.8 GB
3-bit18.1 GB19.2 GB19.8 GB22.4 GB
2-bit13.9 GB14.9 GB15.6 GB18.2 GB
1.6-bit7.6 GB8.7 GB9.4 GB12.0 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run Flint

The best way Flint runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Flint on RTX 4060 report
RTX 5060Too bigPDF: Flint on RTX 5060 report
RTX 3070Too bigPDF: Flint on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Flint on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Flint on RTX 3060 12 GB report
Arc B580Too bigPDF: Flint on Arc B580 report
RTX 4070Too bigPDF: Flint on RTX 4070 report
RTX 4070 SuperToo bigPDF: Flint on RTX 4070 Super report
RTX 5070Too bigPDF: Flint on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: Flint on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: Flint on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: Flint on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: Flint on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: Flint on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: Flint on RTX 4080 Super report
RTX 5070 TiToo bigPDF: Flint on RTX 5070 Ti report
RTX 5080Too bigPDF: Flint on RTX 5080 report
Radeon RX 7900 XTX3-bit136K171 tok/s0.4 sPDF: Flint on Radeon RX 7900 XTX report
RTX 30903-bit136K237 tok/s0.2 sPDF: Flint on RTX 3090 report
RTX 40903-bit136K242 tok/s0.1 sPDF: Flint on RTX 4090 report
Radeon AI PRO R97005-bit112K132 tok/s0.3 sPDF: Flint on Radeon AI PRO R9700 report
RTX 50905-bit112K256 tok/s0.1 sPDF: Flint on RTX 5090 report
2× RTX 30908-bit168K175 tok/s0.1 sPDF: Flint on 2× RTX 3090 report
2× RTX 40908-bit168K181 tok/s0.1 sPDF: Flint on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit192K205 tok/s0.2 sPDF: Flint on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit192K177 tok/s0.3 sPDF: Flint on RTX 6000 Ada report
L40S8-bit192K168 tok/s0.3 sPDF: Flint on L40S report
2× RTX 50908-bit256K228 tok/s0.1 sPDF: Flint on 2× RTX 5090 report
4× RTX 30908-bit256K175 tok/s0.1 sPDF: Flint on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K228 tok/s0.1 sPDF: Flint on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Flint on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: Flint on Mac mini M4 24 GB report
Mac mini M4 32 GB3-bit52K65 tok/s5.6 sPDF: Flint on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: Flint on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB6-bit112K76 tok/s2.8 sPDF: Flint on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit200K67 tok/s2.8 sPDF: Flint on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K89 tok/s1.4 sPDF: Flint on M4 Max 128 GB report
M5 Max 128 GB8-bit256K93 tok/s0.4 sPDF: Flint on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K97 tok/s0.8 sPDF: Flint on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K97 tok/s0.8 sPDF: Flint on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K97 tok/s0.8 sPDF: Flint on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K109 tok/s0.2 sPDF: Flint on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit200K62 tok/s1.1 sPDF: Flint on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K62 tok/s1.1 sPDF: Flint on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K86 tok/s0.8 sPDF: Flint on RTX Spark report
DGX Spark8-bit256K80 tok/s0.9 sPDF: Flint on DGX Spark report
Dual DGX Spark8-bit256K80 tok/s0.5 sPDF: Flint on Dual DGX Spark report
DGX Station GB3008-bit256K303 tok/s0.1 sPDF: Flint on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K228 tok/s0.1 sPDF: Flint on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K228 tok/s0.1 sPDF: Flint on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K228 tok/s0.1 sPDF: Flint on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K237 tok/s0.2 sPDF: Flint on A100 80 GB SXM report
H100 SXM8-bit256K269 tok/s0.1 sPDF: Flint on H100 SXM report
4× H200 NVL8-bit256K287 tok/s0.1 sPDF: Flint on 4× H200 NVL report
8× H200 NVL8-bit256K287 tok/s0.1 sPDF: Flint on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf unsloth/flint-GGUF
vLLM
vllm serve RecursiveSI/flint
MLX (Mac)
mlx_lm.chat --model mlx-community/flint-OptiQ-4bit

Questions

How much VRAM does Flint need?

Flint needs 22.5 GB for its weights at 4-bit, 23.5 GB with an 8K context and 24.2 GB with 32K. At 8-bit, 40.4 GB.

What is the minimum VRAM for Flint?

About 19.2 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the Radeon RX 7900 XTX.

How big is Flint?

39.8B parameters (1.7B active per token): 74.1 GB of weights at 16-bit, 22.5 GB at 4-bit.

Can Flint run on an RTX 4090?

Yes. At 3-bit it holds 136K of context and writes about 242 tokens a second.

More models