Recursive Flint: VRAM requirements by quant
Key facts
Recursive Flint needs about 24.5 GB at standard quality (Q4, estimate).
Recursive Flint is a mid-size AI model you can download and run on your own computer. It is a mixture of experts: all of it has to fit in memory, but it uses only a small part for each word. It needs about 24 GB of memory at standard quality.
Try Recursive Flint free in your browser, no login, no install
LIVE IN BROWSER · $0 · NO LOGIN
Test Recursive Flint (40B MoE) free in your browser before downloading 23.39 GB
Open full 8-model playground ↗You can test Recursive Flint (recursive/flint-v1.0:free, 39.79B total parameters with 2.5B active per token; 262,144 max context; Apache-2.0, our reading of the licence, not legal advice) free in your browser with no login and 0 GB of local VRAM. Send one coding or reasoning prompt in the box below or open the full side-by-side chat at https://localllmrig.com/ to compare Recursive Flint against Gemma 4 31B, Gemma 4 26B A4B, and Laguna S 2.1 before downloading the local GGUF (24.46 GB total estimate at Q4_K_M = 23.39 GB measured weights + 0.27 GB KV cache at 8,192 tokens + 0.80 GB overhead, or 19.56 GB total at UD-Q3_K_XL to fit a single 24 GB GPU).
Test Recursive Flint vs 9 other free open models on localLLMrig →https://localllmrig.com/can-i-run/flint
Click a starter prompt below or type your own question to watch Recursive Flint stream side by side against Gemma 4 31B right on this page:
Memory needed
About 24 GB
At standard quality, with room for a conversation of about 12 pages.
Get a 1-line alert when a faster quant for Recursive Flint drops — or when GPUs that run it drop >10%
Get a 1-line alert when a new open-weight model fits 32 GB VRAM (24.5 GB at Q4 for Recursive Flint) or when used RTX 3090 / 4090 & Mac Studio prices drop >10% — weekly, zero spam.
MoE 35B-A3B vs dense 27B on 24 GB: Qwen3.6-35B-A3B vs Qwen3.8-27B.
Machines that run Recursive Flint
Each of these holds the whole model in memory. We do not estimate a speed for this kind of model (a mixture of experts), so these are simply the cheapest.
EU shops · Shown to visitors in Europe.US shops · Shown to visitors outside Europe.
If you buy through these links we may earn a commission. It doesn't change our ranking. How shop links work
Mac Studio M4 Max 36GB
36 GB shared memory · second-hand
$2,116€1,890usedunverified
Machine details →Mac mini M4 Pro 64GB
64 GB shared memory · second-hand
$2,116€1,890usedunverified
Machine details →Mac Studio M5 Max 36GB
36 GB shared memory
$2,802€3,029new
Machine details →Mac mini M5 Pro 64GB
64 GB shared memory
$2,961€3,200estimate
Machine details →Mac Studio M4 Max 128GB
128 GB shared memory · second-hand
$3,683€3,290usedunverified
Machine details →
Where our price is for a used card or Mac, these shops mostly sell new ones, so their price can be higher.
As an Amazon Associate I earn from qualifying purchases.
Als Amazon-Partner verdiene ich an qualifizierten Verkäufen.
Als Amazon-Partner verdien ik aan in aanmerking komende aankopen.
About these prices
Each price is the amount we last wrote down for that machine, with its date on the machine page. "Estimate" is our own figure: for a complete PC we add our estimate for the other parts to the graphics card price, and some Mac prices are estimates, not a checked shop listing. "Used" is a second-hand price. "Unverified" means we have no checked shop listing for that exact amount. The shop buttons open a search for that machine at each shop, so you see today's price there. Prices excl. shipping unless marked. Euro prices incl. VAT where the seller charges it. Dollar prices are converted from our euro prices without the 21% EU VAT, before US sales tax.
Or rent one by the hour: RTX A6000 at Vast.ai holds it, for $0.40 an hour (median of live offers, 21 Oct 2026). Compare GPU cloud prices
Speeds and memory are our estimate from the hardware specs, not a measured test. Prices are dated and may change. How we calculate
Conversation memory for Recursive Flint, our estimate: 35 of 49 layers keep a 513-token sliding window and 14 layers cache the full context. At 8,192 tokens that cache is about 0.27 GB, at 32,768 tokens about 0.98 GB, at 131,072 tokens about 3.79 GB, and at 262,144 tokens about 7.55 GB. llama.cpp needs a build that includes Flint (b11803 or newer; leave --swa-full unset). Ollama: ollama run flint.
Memory at each quality level
A lower quality level makes the model smaller, so it fits on a smaller machine. Standard (Q4) is what most people run.
| Quality | Memory needed |
|---|---|
| Q4 · standard (smallest that keeps quality) | 24.5 GB |
| Q5 · a bit sharper, a bit bigger | 29.3 GB |
| Q8 · near full quality, almost twice Q4 | 43.6 GB |
| FP16 · full quality, over three times Q4 | 80.7 GB |
With a conversation of 8,192 tokens · about 12 pages of text.
Sources and method
At standard quality: 23.4 GB for the model itself (a measured file), 0.3 GB for a conversation of 8,192 tokens, and 0.8 GB for the software.
Size source: huggingface.co/RecursiveSI/flint. 39.79B parameters in total, 2.5B used for each word. The memory counts the total.
Recursive's coding model for agents that read code, fix bugs and write tests, under Apache 2.0. Only about 2.5B of its 40B parameters are used per word, so it does far less work per word than a dense 40B model. The 4-bit downloads are about 22 to 26 GB before any conversation memory.
Licence
Commercial use: yes
Commercial use allowed. Keep the licence and copyright notice when you share the weights.
Licence: Apache-2.0 · checked 23 Oct 2026
Our reading of the licence, not legal advice.
Will Recursive Flint run on my graphics card?
Common questions
- How much memory does Recursive Flint need?
- About 24.5 GB at standard quality (Q4) with a conversation of 8,192 tokens: 23.4 GB for the model itself, 0.3 GB for the conversation and 0.8 GB for the software. The model size is a measured file; the conversation and software parts are our estimate.
- Does Recursive Flint fit on a single 24 GB graphics card?
- Not at 4-bit Q4_K_M (24.46 GB total estimate vs 23.4 GB usable on a 24 GB card). It fits on a single 24 GB card at 3-bit UD-Q3_K_XL (18.49 GB measured GGUF from unsloth/flint-GGUF, 19.56 GB total estimate at 8,192 context, leaving about 3.8 GB to spare).
- Are the speeds measured?
- We do not show a speed for this model. It is a mixture-of-experts model, and our estimate would not be reliable for that design.
- Where can I test Recursive Flint online for free before downloading 23.39 GB?
- You can test Recursive Flint (recursive/flint-v1.0:free), 39.79B total, 2.5B active, live in your browser for free at https://localllmrig.com/ or from https://localllmrig.com/can-i-run/flint, with no account and no graphics card. It is ticked by default on the home page. In Choose LLMs under Hardware $1,500–5,000, tick it alongside other free models, such as Gemma 4 31B, Gemma 4 26B A4B, NVIDIA Nemotron 3.5 Lightning, NVIDIA Nemotron 3.5 Super VL, and Laguna S 2.1, and compare code and reasoning on your own prompts (5 free questions every 10 minutes).
- Can Recursive Flint run on a 24 GB or 16 GB GPU?
- Q4_K_M (23.39 GB measured GGUF, 24.5 GB total estimate) at 8,192 tokens is 1.1 GB over a 24 GB card's 23.4 GB usable memory. The standard file fits 3 of the 10 graphics-card setups on https://localllmrig.com/can-i-run/flint: GeForce RTX 5090 32GB (€7,367 / $6,816); Dual RTX 3090, 128 GB RAM (€5,452 / $5,528, price unverified); Dual RTX 4090, 128 GB RAM (€7,330 / $7,631, price unverified). IQ4_XS (22.45 GB measured GGUF, 23.5 GB total estimate) is still 0.1 GB over a 24 GB card. UD-Q3_K_XL (18.49 GB measured GGUF, 19.6 GB total estimate) fits a single 24 GB card such as GeForce RTX 3090 24GB (€1,246 / $1,395), with 3.8 GB to spare. On a 16 GB card (15.4 GB usable), UD-Q2_K_XL (13.56 GB measured GGUF, 14.6 GB total estimate) fits, and so does UD-IQ1_M (11.35 GB measured GGUF, 12.4 GB total estimate).
- How does Recursive Flint keep its 262,144-token context window so light in memory?
- Recursive Flint has 49 layers (4 key-value heads, head size 128). 35 of those layers use a 513-token sliding window, and 14 layers keep the full context. Conversation memory is 0.27 GB at 8,192 tokens (0.3 GB rounded) and 7.55 GB across the full 262,144-token context (about 393 pages), our estimate. The same figures are on https://localllmrig.com/can-i-run/flint.
- Which llama.cpp version or Ollama command is required to run Recursive Flint locally?
- llama.cpp needs a build that includes Flint (b11803 or newer; leave --swa-full unset). Ollama: ollama run flint. The measured Q4_K_M file in this catalog is 23.39 GB. A 24 GB card holds UD-Q3_K_XL (18.49 GB measured GGUF, 19.6 GB total estimate); the standard file needs 32 GB of graphics memory or a Mac that leaves at least 24.5 GB usable.
- Which Macs, Ryzen AI Max+ 395 mini PCs, or DGX Spark machines run Recursive Flint best?
- At Q4_K_M (23.39 GB measured GGUF, 24.5 GB total estimate), a 32 GB Mac leaves about 21.3 GB usable, which is 3.1 GB short of the standard file. These measured files do fit Mac mini M6 32GB (21.3 GB usable, €1,530 / $1,416, price unverified): UD-Q3_K_XL (18.49 GB measured GGUF, 19.6 GB total estimate), UD-IQ3_XXS (14.54 GB measured GGUF, 15.6 GB total estimate), UD-Q2_K_XL (13.56 GB measured GGUF, 14.6 GB total estimate), UD-IQ1_M (11.35 GB measured GGUF, 12.4 GB total estimate). Complete machines on https://localllmrig.com/can-i-run/flint that hold the standard file, cheapest first: Mac mini M4 Pro 64GB (48.0 GB usable, €1,890 / $2,116, price unverified); GMKtec EVO-X2 (Ryzen AI Max+ 395) 64GB (48.0 GB usable, €2,199 / $2,035); Mac Studio M5 Max 36GB (27.0 GB usable, €3,029 / $2,802); Mac mini M5 Pro 64GB (48.0 GB usable, €3,200 / $2,961, price unverified); GMKtec EVO-X2 (Ryzen AI Max+ 395) 128GB (96.0 GB usable, €3,800 / $3,516); GEEKOM A9 Mega (Ryzen AI Max+ 395) 128GB (96.0 GB usable, €3,999 / $3,700); Beelink GTR9 Pro (Ryzen AI Max+ 395) 128GB (96.0 GB usable, €4,619 / $4,274); Mac Studio M5 Max 128GB (96.0 GB usable, €6,000 / $5,551, price unverified); Mac Studio M5 Ultra 96GB (72.0 GB usable, €6,649 / $6,152); Mac Studio M5 Ultra 256GB (192.0 GB usable, €12,500 / $11,565, price unverified). Larger files: Q5_K_M (30.81 GB measured GGUF, 31.9 GB total estimate) on Mac mini M4 Pro 64GB (48.0 GB usable); Q6_K (35.2 GB measured GGUF, 36.3 GB total estimate) on Mac mini M4 Pro 64GB (48.0 GB usable); Q8_0 (42.35 GB measured GGUF, 43.4 GB total estimate) on Mac mini M4 Pro 64GB (48.0 GB usable); BF16 (79.63 GB measured GGUF, 80.7 GB total estimate) on GMKtec EVO-X2 (Ryzen AI Max+ 395) 128GB (96.0 GB usable). We don't give a speed for this mixture-of-experts model yet.
- How fast does Recursive Flint generate tokens in real hardware benchmarks versus our formula?
- We don't give a speed for this model yet. This model only uses part of itself for each word (2.5B active of 39.79B total, a mixture of experts), and our speed estimate does not handle that yet. Recursive-reported (announcement, not a measurement on this site). With Recursive's own inference engine, Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 tokens per second on a 64GB NVIDIA RTX Spark at short contexts. Source: https://recursive.com/updates/introducing-recursive-flint. An external write-up we did not run (InferenceKey) reports about 210 tokens per second on an RTX 5090 at 64K context and about 167 on an RTX 4090 at 16K, and says 32K ran out of memory on the 4090. Those figures are community measurements, not a localLLMrig measurement. The hardware table is on https://localllmrig.com/can-i-run/flint.
- How do Recursive Flint, Llama 3.3 70B and DeepSeek R1 Distill 70B compare with Claude and GPT?
- Want to see how local open models (Recursive Flint, Llama 3.3 70B, DeepSeek R1 Distill 70B) stack up against Claude Sonnet 5.5 and GPT-6.1 Sol before buying hardware? Run them side by side at https://localllmrig.com/ and read https://localllmrig.com/guides/compare-claude-and-gpt-with-local-llms.
- Can I try Recursive Flint online for free?
- Yes. Recursive Flint is on the free chat on the home page. Pick it there, ask a question, and read the answer in your browser. No login and no install.
- Do I need a GPU to test it?
- No. The free chat runs in your browser. It does not download the model and it does not use a graphics card on your computer. A graphics card only matters if you run the model yourself.
- Do I need an account to try Recursive Flint?
- No. The free chat needs no account and no login.