Nanbeige4.3-33B: VRAM requirements by quant
Key facts
Nanbeige4.3-33B needs about 23.3 GB at standard quality (Q4, estimate).
Checked . bartowski Q4_K_M GGUF.
Nanbeige4.3-33B is a mid-size AI model you can download and run on your own computer. It needs about 23 GB of memory at standard quality.
LIVE IN BROWSER · $0 · NO LOGIN
Test 10 open LLMs free in your browser before downloading 21 GB for Nanbeige4.3-33B
Open full 8-model playground ↗Before downloading 20.79 GB of Q4_K_M weights for Nanbeige4.3-33B (33.12B dense parameters, 524,288 max context, Apache-2.0; author-reported Terminal-Bench 80.9 and DeepSWE 65.5, not independently verified on this site), you can test 10 open LLMs free in your browser at https://localllmrig.com/ with no login and 0 GB of local VRAM. The prompts below run Recursive Flint and Gemma 4 31B. Nanbeige4.3-33B itself is not one of the free models, so trying it in the chat needs a pass once a provider lists it.
Test 10 open LLMs free in your browser →https://localllmrig.com/can-i-run/nanbeige-4-3-33b
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 23 GB
At standard quality, with room for a conversation of about 12 pages.
Get a 1-line alert when a faster quant for Nanbeige4.3-33B drops — or when GPUs that run it drop >10%
Get a 1-line alert when a new open-weight model fits 24 GB VRAM (23.3 GB at Q4 for Nanbeige4.3-33B) or when used RTX 3090 / 4090 & Mac Studio prices drop >10% — weekly, zero spam.
Machines that run Nanbeige4.3-33B
Each of these holds the whole model in memory. Speed is in tokens per second (tok/s); from about 8 it keeps up with your reading.
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
Comfortable~8.7 tok/s$2,116€1,890usedunverified
Machine details →Mac mini M4 Pro 64GB
64 GB shared memory · second-hand
Slow~5.8 tok/s$2,116€1,890usedunverified
Machine details →Used PC with RTX 3090 24GB
24 GB graphics memory · second-hand
Fast~32 tok/s$2,413€2,346estimateused parts
Machine details →The buttons look up the graphics card, the part of this PC worth comparing.
PC with Radeon RX 7900 XTX 24GB
24 GB graphics memory · second-hand
Fast~32 tok/s$2,422€2,354estimateused parts
Machine details →The buttons look up the graphics card, the part of this PC worth comparing.
PC with RTX 5090 32GB
32 GB graphics memory
Fast~61 tok/s$7,834€8,467estimate
Machine details →The buttons look up the graphics card, the part of this PC worth comparing.
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
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) | 23.3 GB |
| Q5 · a bit sharper, a bit bigger | 26.1 GB |
| Q8 · near full quality, almost twice Q4 | 38.0 GB |
| FP16 · full quality, over three times Q4 | 68.8 GB |
With a conversation of 8,192 tokens · about 12 pages of text.
Sources and method
At standard quality: 20.8 GB for the model itself (a measured file), 1.7 GB for a conversation of 8,192 tokens, and 0.8 GB for the software.
Size source: huggingface.co/Nanbeige/Nanbeige4.3-33B. The published size is rounded, so ours is an estimate.
Dense coder and reasoning model, 33.12B parameters, Apache-2.0 on the model-card metadata (the repo had no LICENSE file on 26 Oct 2026). Config: 52 layers, 8 KV heads, head dim 128, context 524,288. The parameter formula is about 19.9 GB at Q4, 35.4 GB at Q8 and 66.2 GB at FP16. Bartowski's GGUF listing on 26 Oct 2026 measured Q4_K_M at 20.79 GB, Q8_0 at 35.20 GB and the two BF16 shards at 66.24 GB together. IQ2_M on the same listing measured 12.19 GB on 27 Oct 2026 (12,187,167,872 bytes). At 8,192 tokens our estimate adds about 1.74 GB of conversation memory and 0.80 GB of working space, so the Q4_K_M total is about 23.3 GB. That fits a 24 GB GPU (an RTX 3090 or RTX 4090 leaves about 23.4 GB usable). A 32 GB Mac leaves about 21.3 GB for the GPU, so the same Q4 total is over that default share; IQ4_XS at 18.64 GB is the file that lands inside it. The IQ2_M total is about 14.7 GB, which fits a 16 GB card and misses a 12 GB card. Q8_0 at 35.20 GB fits 48 GB of VRAM or 64 GB of unified memory. FP16 at 66.24 GB fits 96 GB or 128 GB of unified memory, or two 48 GB GPUs. The Terminal-Bench 80.9 and DeepSWE 65.5 figures are the author's, and are not yet independently verified.
Licence
Commercial use: yes
Commercial use allowed. Keep the licence and copyright notice when you share the weights.
Licence: Apache-2.0 · checked 26 Oct 2026
Our reading of the licence, not legal advice.
Will Nanbeige4.3-33B run on my graphics card?
Builder interview
Xijing (Thomas) Wang is testing Nanbeige4.3-33B on a MacBook Pro M5 Max with 128 GB unified memory (about 107 GB of that is GPU-usable), alongside his daily Qwen3.8-27B at 5-bit on oMLX. Self-reported, 26 October 2026. Read Xijing (Thomas) Wang's interview.
Common questions
- How much memory does Nanbeige4.3-33B need?
- About 23.3 GB at standard quality (Q4) with a conversation of 8,192 tokens: 20.8 GB for the model itself, 1.7 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 Nanbeige4.3-33B fit on a single 24 GB graphics card?
- Yes, by our estimate. It needs about 23.3 GB, and we count 23.4 GB of a 24 GB card as usable.
- Are the speeds measured?
- No. They are our estimate from how fast each machine's memory is. Real speed also depends on the software and its settings.
- Can I run Nanbeige4.3-33B on an RTX 3090 or RTX 4090 (24 GB VRAM)?
- Yes, by our estimate, and it is a tight fit. At 8,192 tokens, 20.79 GB measured GGUF + 1.74 GB conversation memory at 8,192 tokens + 0.80 GB working space = 23.33 GB / 23.3 GB total estimate (52 layers, 8 KV heads, head dim 128). A 24 GB card leaves 23.4 GB usable, so Q4_K_M fits with about 0.07 GB to spare. 6 of the 10 GPU setups on https://localllmrig.com/can-i-run/nanbeige-4-3-33b hold the standard file: GeForce RTX 3090 24GB (€1,246 / $1,395, about 32 tok/s, our estimate); GeForce RTX 4090 24GB (€2,185 / $2,446, about 35 tok/s, our estimate); GeForce RTX 5090 32GB (€7,367 / $6,816, about 61 tok/s, our estimate); Radeon RX 7900 XTX 24GB (€1,254 / $1,404, about 32 tok/s, our estimate); Dual RTX 3090, 128 GB RAM (€5,452 / $5,528, price unverified, about 35 tok/s, our estimate); Dual RTX 4090, 128 GB RAM (€7,330 / $7,631, price unverified, about 38 tok/s, our estimate). GeForce RTX 3090 24GB, GeForce RTX 4090 24GB, and Radeon RX 7900 XTX 24GB each leave that 23.4 GB usable. For a longer conversation on one 24 GB card, IQ4_XS (18.64 GB measured GGUF, 21.2 GB total estimate) frees 2.15 GB, or quantize the conversation memory (-ctk q8_0 -ctv q8_0). Q5_K_M (25.07 GB measured GGUF, 27.6 GB total estimate) is 4.2 GB over a 24 GB card and needs a 32 GB GPU (GeForce RTX 5090 32GB, 31.4 GB usable, about 61 tok/s, our estimate), two 24 GB cards, or a 64 GB Mac or Ryzen AI Max+ 395 mini PC.
- Can a 16 GB or 12 GB GPU (RTX 4060 Ti 16GB, RTX 5080 16GB, or RTX 3060 12GB) run Nanbeige4.3-33B?
- Not at standard 4-bit quality (Q4_K_M, 23.3 GB total estimate) or IQ4_XS (21.2 GB total estimate). A 16 GB card (GeForce RTX 4060 Ti 16GB at €899 / $832, or GeForce RTX 5080 16GB) leaves 15.4 GB usable by our estimate. Q3_K_M (15.92 GB measured GGUF, 18.5 GB total estimate), IQ3_XS (15.21 GB measured GGUF, 17.8 GB total estimate), IQ3_XXS (14.49 GB measured GGUF, 17.0 GB total estimate) all exceed 15.4 GB at 8,192 tokens. To fit entirely inside a 16 GB card, step down to IQ2_M (12.19 GB measured GGUF, 14.7 GB total estimate) or IQ2_XXS (10.07 GB measured GGUF, 12.6 GB total estimate), which trades away coding and reasoning accuracy. GeForce RTX 3060 12GB leaves 11.4 GB usable (€261 / $292) and cannot hold IQ2_XXS. Budget PC with RTX 3060 12GB runs Q4_K_M by spilling into normal memory at about 1.8 tok/s, our estimate. Qwen3.8-27B (https://localllmrig.com/can-i-run/qwen3.8-27b) at UD-IQ3_XXS (10.93 GB measured GGUF, 12.3 GB total estimate) or Gemma 4 26B A4B (https://localllmrig.com/can-i-run/gemma-4-26b-a4b) at UD-IQ4_XS (13.6 GB measured GGUF, 14.8 GB total estimate).
- Which Mac or Ryzen AI Max+ 395 mini PC runs Nanbeige4.3-33B best, and why is a 32 GB Mac tight at Q4_K_M?
- Mac mini M6 32GB leaves 21.3 GB usable of 32 GB (€1,530 / $1,416, price unverified), 2.0 GB short of the 23.3 GB Q4_K_M total estimate at 8,192 tokens. IQ4_XS (18.64 GB measured GGUF, 21.2 GB total estimate) is the file that fits a 32 GB Mac. Q4_K_M (20.79 GB measured GGUF, 23.3 GB total estimate) does not. Mac Studio M4 Max 36GB (27.0 GB usable, €1,890 / $2,116, price unverified, about 8.7 tok/s, our estimate) and Mac Studio M5 Max 36GB (27.0 GB usable, €3,029 / $2,802, about 9.8 tok/s, our estimate) hold Q4_K_M at 23.3 GB. GMKtec EVO-X2 (Ryzen AI Max+ 395) 64GB leaves 48.0 GB for graphics (€2,199 / $2,035) and holds Q4_K_M (23.3 GB, about 9.1 tok/s, our estimate). Mac Studio M5 Max 128GB leaves 96.0 GB usable (about 13.0 tok/s, our estimate). Mac Studio M5 Ultra 96GB leaves 72.0 GB usable and Mac Studio M5 Ultra 256GB leaves 192.0 GB usable (about 25 tok/s, our estimate). Q5_K_M (27.6 GB total), Q6_K (31.6 GB total), Q8_0 (37.7 GB total) fit 64 GB unified memory (48.0 GB usable). BF16 (66.24 GB measured GGUF, 68.8 GB total estimate) fits 96 GB or 128 GB unified memory. A 128 GB M5 Max write-up is at https://localllmrig.com/interviews/thomas-wang-m5-max-128gb-omlx. Ryzen AI Max+ 395 is the chip in GMKtec EVO-X2 (Ryzen AI Max+ 395) 64GB.
- Nanbeige4.3-33B vs Qwen3.8-27B: how do their VRAM requirements and 24 GB GPU fit compare?
- Both are dense Apache-2.0 models that fit a single 24 GB GPU (23.4 GB usable), but Qwen3.8-27B needs 5.53 GB less VRAM at standard 4-bit quality and 3.25× less conversation memory per token, by our estimate. Nanbeige4.3-33B is 33.12B parameters with 52 full-attention layers, 23.33 GB / 23.3 GB at Q4_K_M (20.79 GB measured GGUF + 1.74 GB conversation memory + 0.80 GB working space), leaving 0.07 GB on a 24 GB card (about 32 tok/s, our estimate on an RTX 3090, about 35 tok/s, our estimate on an RTX 4090). Qwen3.8-27B is 27.78B parameters with 48 linear-attention layers and 16 full-attention layers, 17.80 GB at UD-Q4_K_M (0.54 GB versus 1.74 GB of conversation memory at 8,192 tokens). UD-Q5_K_XL is 22.2 GB total. UD-Q6_K is 23.3 GB total. At 32,768 tokens, Qwen3.8-27B UD-Q4_K_M is 19.41 GB and Nanbeige4.3-33B is 28.57 GB. Qwen3.8-27B is at https://localllmrig.com/can-i-run/qwen3.8-27b.
- Where can I test open coding and reasoning LLMs free in my browser before downloading 20.79 GB for Nanbeige4.3-33B?
- Nanbeige4.3-33B is not one of the 10 free models. Before downloading 20.79 GB of Q4_K_M weights you can test 10 open-weight models side by side for free in your browser at https://localllmrig.com/ or from https://localllmrig.com/can-i-run/nanbeige-4-3-33b, with no login and 0 GB of local VRAM. Compare Recursive Flint (39.79B total, 2.5B active, https://localllmrig.com/can-i-run/flint), Gemma 4 31B (31.27B dense, https://localllmrig.com/can-i-run/gemma-4-31b), and Gemma 4 26B A4B (https://localllmrig.com/can-i-run/gemma-4-26b-a4b) on your own coding and reasoning prompts (5 free questions every 10 minutes). Trying Nanbeige4.3-33B itself in the chat needs a pass once a provider lists it.
- What licence does Nanbeige4.3-33B use, and how do I run it locally with llama.cpp, Ollama, or vLLM on 24 GB?
- The Nanbeige/Nanbeige4.3-33B model-card metadata lists Apache-2.0 (the repository had no standalone LICENSE file on 26 Oct 2026; our reading of the licence, not legal advice). It uses standard LlamaForCausalLM (52 layers, 8 KV heads, head dim 128, 524,288 max context) and runs in current llama.cpp, Ollama, vLLM, SGLang, and oMLX builds. On a 24 GB GPU (GeForce RTX 3090 24GB, GeForce RTX 4090 24GB, or Radeon RX 7900 XTX 24GB), run Nanbeige_Nanbeige4.3-33B-Q4_K_M.gguf (20.79 GB) in llama.cpp with -ngl 99 -fa -c 8192, or add -ctk q8_0 -ctv q8_0 (or use IQ4_XS at 18.64 GB) for a longer context. File sizes are on https://localllmrig.com/can-i-run/nanbeige-4-3-33b.