8 GB or 16 GB GPU for Flint 40B? Both run it; one reads faster

Short answer

You don’t need an expensive graphics card to run Flint 40B. An 8 GB RTX 4060 and a 16 GB RX 7800 XT both ran it in LM Studio, with part of the model in system RAM; graphics memory plus RAM must hold about 24 GB at 8K context. The 16 GB card wrote 20.4 tokens per second on a long prompt, the 8 GB card 14.6.

The 8 GB card isn’t slower at thinking, it’s slower at reading

I ran the same Flint 40B file on two desktop PCs on November 2 and 3, 2026, with my own benchmark tool, SwarmBench. My RTX 4060 PC has an NVIDIA RTX 4060 with 8 GB of graphics memory (VRAM). The other, a second desktop, has an AMD RX 7800 XT with 16 GB.

Both cards ran the model. Both passed nearly the same number of quality tasks. The difference was speed, and most of it showed up when the model read a long prompt, not when it wrote the answer.

The reason is simple. The model file is bigger than either card’s memory, so part of it always sits in system RAM, which is slower for the graphics card to reach. The 16 GB card holds more of the model, so less of the work waits on system RAM. The model is the same; the extra VRAM buys speed, not smarts.

A few terms:

• Token. A piece of a word. In English, 100 tokens are about 75 words.
• Prompt reading. Before the model writes, it reads your whole prompt. Our reading vs writing guide explains why the two speeds differ.
• Context size. How much text the model can keep in view at one time: the prompt, its thinking and the answer. You set it when you load the model. A bigger context needs more memory, so less of the model fits on the graphics card.

What the context size does to reading speed

This is the result that matters most if you plan to run long documents or long coding sessions. It shows prompt reading speed with a 5,000-token prompt.

What this means in plain terms:

• On the 16 GB card, context up to 128K is free. Reading speed showed no change from 8K to 128K. Only the maximum, 262K, slowed it down, from 321 to 174 tokens per second.
• Writing speed did not change. At 262K, the RX 7800 XT still wrote 20.4 tokens per second with the long prompt. A big context makes you wait longer for the first word, not for the rest of the answer.
• The 8 GB card loses more. At 262K, the RTX 4060 read the prompt at 102 tokens per second, against 277 at 8K. The first word took 49 s, not 18.0 s.

So on a 16 GB card, 128K context cost nothing in these tests. Use the maximum only when you need the room, for example for a very long document or very long thinking. On an 8 GB card, keep the context as small as your task allows.

Writing speed

With a long prompt (5,000 tokens), the RX 7800 XT wrote 20.4 tokens per second and the RTX 4060 wrote 14.6. That is 40% faster on the 16 GB card. With a short prompt, the RX 7800 XT was 8% faster.

In these tests, the gap was bigger with the long prompt than with the short one.

Quality: almost the same

Every score comes from a program, not an AI judge. The answer limit is how many tokens Flint may use to think and answer; an answer that hits it fails.

• Most failures were not wrong answers. Flint thought for too long and hit the answer limit before it wrote the answer. With the larger limit, almost all of these tasks passed. More room to think helped more than more VRAM.
• The one rule failure on the RTX 4060 was the “no letter e” task. Flint repeated the same thoughts for 77 minutes and hit the limit. On the RX 7800 XT, the same task passed in 66 seconds. NVIDIA and AMD cards do the math in slightly different ways, so the model can take a different path. One task is not enough to say one card gives better answers.

Small models are not a shortcut. As a side note, the same tests ran a different model, Qwen3.5 2B, on a different machine, a Mac mini M4 with 16 GB. It wrote 54 to 59 tokens per second, but passed only 6 of 16 code tasks with thinking off. It is fast because it is small. It is not part of the comparison above.

Cost

The table uses the 3-month median new price in the US from gpuprix.com, read on November 4, 2026. This is not today’s price. Prices for these cards move fast: on November 3, the lowest in-stock RTX 4060 was $460, and one day later it was $600.

Lower is less money per unit of speed. With long prompts, the two cards come out almost the same: the RX 7800 XT cost about 38% more at these prices and wrote 40% faster. With short prompts, the RTX 4060 gave more speed per dollar. Check today’s prices before you buy; a few days can change the picture.

The card is not the only cost. You also need enough system RAM, as the next section explains.

How much memory Flint needs

The download size is not the memory requirement. The file used here is the Q4_K_M version, 24.2 GB. When you load it, the model also needs memory for the context. LM Studio estimates about 24 GB at 8K context and 36.5 GB at the maximum context (262K). Graphics memory and system RAM together must hold that.

Neither card holds the whole model. On both PCs, LM Studio put part of the model on the graphics card and kept the rest in system RAM. This is called RAM offload. It is not the same as streaming a model from an SSD, and it is not a cluster of several machines: it is one PC using its own graphics memory and its own RAM.

The two PCs had 64 GB (RTX 4060) and 32 GB (RX 7800 XT) of DDR5-6000 RAM. To check what fits on your own machine, use My rig.

About Flint

Flint 40B is a model from Recursive SI. Its Hugging Face repo was created on October 21, 2026.

• 39.8 billion parameters in total (about 40B), about 2.5 billion active for each token.
• A mixture-of-experts (MoE) model: it uses only a small part of its weights for each token, which is why it runs on home hardware at all.
• Up to 262,144 tokens of context.
• Apache 2.0 license.

Benchmark record

Measured by me with SwarmBench, my own benchmark tool; first-party, self-measured; not RigPerDollar estimates.

Limits

• One run per quality task. At temperature 0, a different graphics card can still give a different answer. A 1-task gap (17 of 18 against 18 of 18) may be luck. A stronger test runs each task 3 to 5 times.
• The machines differ in more than the graphics card. The CPUs (i5-14600K, i7-13700K) and the RAM sizes (64 GB, 32 GB) are different. The RAM speed is the same.
• No power data. I didn’t measure the watts, so there is no electricity cost here.
• LM Studio chose the GPU offload. I didn’t set the share of the model on the graphics card by hand. Manual settings may make either machine faster.
• The AMD PC’s operating system, driver and runtime backend were not recorded. Vulkan or ROCm can change the speed.
• The RTX 4060 context sweep covers only 8K and 262K. The “no change up to 128K” result is for the RX 7800 XT only.
• LM Studio version: Unknown on both machines.
• Temperature. Recursive’s model card recommends temperature 1.0. These tests used 0, so scores may differ at the recommended settings.