Nanbeige4.3 33B VRAM requirements: can you run it locally?

Nanbeige4.3 33B is a dense open-weight model from Nanbeige, built for coding agents and office-style agent tasks. Its Hugging Face repo was created on October 24, 2026, it is Apache 2.0 licensed, and Hugging Face lists 33,118,977,024 parameters.^[1] Nanbeige announced it on X on October 25, 2026.^[13] Its config.json uses the standard Llama architecture with 52 layers and sets a maximum of 524,288 positions (512K tokens), which the model card also states.^[3][2]

The card compares it with Qwen3.8 27B, Muse-Glimmer-30B and Gemma 4 31B and reports, for example, 80.9 on Terminal-Bench 2.1 (Qwen3.8 27B: 73.0) and 60.3 on SWE-Bench Pro (Qwen3.8 27B: 61.7). These are developer-reported scores.^[2]

Memory needed at each quantization

Quantization stores each weight in fewer bits. Smaller files fit on smaller GPUs, at some cost in quality, and the cost grows as you go below about 4 bits. File sizes below are the actual GGUF downloads on Hugging Face (1 GB = 1,000,000,000 bytes).^[4]

Context length adds memory (the KV cache)

This is where Nanbeige4.3 33B differs most from Qwen3.8 27B. It is a plain dense model, so all 52 layers keep a key/value cache that grows with every token.^[3] With 8 KV heads and a head size of 128, an FP16 cache takes 2 × 52 × 8 × 128 × 2 bytes = 212,992 bytes per token, more than three times Qwen3.8 27B's 65,536:

So the 512K context is real but expensive: at FP16 the full 524,288 tokens need about 111.7 GB of cache on top of the weights. On a single 24 GB card with Q4_K_M weights, the 3.2 GB left after the 20.8 GB of weights holds at most about 15,000 tokens of FP16 cache, before any engine overhead. An 8-bit or 4-bit KV cache (llama.cpp's --cache-type-k / --cache-type-v options) halves or quarters these numbers, with some quality cost.

Which GPUs and Macs it fits on

Each cell compares the weights plus an 8K-token KV cache with the device's memory.^{[5][6][7][8][9][10][11]}

Column headers show weights + 8,192-token KV cache.

How to read it: "Fits" means weights + 8K cache use under 90% of memory. "Tight" means 90–100%: it may load, but the engine's own working memory or a longer context can push it over. "No" means it won't fit on the GPU alone. The 90% line is our own rule of thumb, not a vendor figure. On a Mac, macOS also needs some of the unified memory, so treat Mac rows as optimistic.

How to run it

With llama.cpp or LM Studio, download a GGUF file from the quant list above, for example Nanbeige_Nanbeige4.3-33B-Q4_K_S.gguf (19.7 GB) or Nanbeige_Nanbeige4.3-33B-IQ4_XS.gguf (18.6 GB) for a 24 GB card, or Nanbeige_Nanbeige4.3-33B-IQ2_M.gguf (12.2 GB) for a 16 GB card.^[4] For vLLM or SGLang, the model card's examples use tensor parallelism across 4 GPUs (--tensor-parallel-size 4 / --tp-size 4) with the BF16 weights.^[2] When we checked, Ollama's library had no official Nanbeige page.

If the model is bigger than your VRAM, these tools can keep some layers in system RAM, but every token then has to wait on the slower memory, so expect much lower speed.

What people report (third-party, not verified by us): one X user ran a 4-bit-per-weight EXL3 build on a single RTX 5090 at "40+ tok/s".^[14]

Cheapest GPU to rent for Nanbeige4.3 33B

A 24 GB GPU can run a 4-bit build at a short context; a 32 GB or 48 GB card gives you room for longer contexts. On-demand Secure Cloud prices listed on Runpod's pricing page:^[12]

Prices checked on 2026-10-26. They change often, so confirm on the provider's page. Rent on Runpod Referral link

Frequently asked questions

Can I run Nanbeige4.3 33B on an RTX 3090 or RTX 4090?

Yes, at 4 bits with a short context. Q4_K_M (20.8 GB) plus an 8K cache is about 22.5 GB, or 94% of a 24 GB card, which is tight.^[4][3] Q4_K_S (19.7 GB, about 90% with 8K) or IQ4_XS (18.6 GB, about 85%) are safer picks. For long contexts, use a quantized KV cache.

Can I run Nanbeige4.3 33B on a 16 GB or 12 GB GPU?

On 16 GB, such as an RTX 5080, IQ2_M (12.2 GB) fits with an 8K context; Q2_K (13.5 GB) is tight and the 3-bit files do not fit with the cache.^[4][6] On 12 GB, such as an RTX 3060, even IQ2_XXS (10.1 GB) is tight once an 8K cache is added, so you would need a shorter context or partial offload to system RAM.^[4][5]

How much VRAM does the full 512K context need?

At FP16, the KV cache for 524,288 tokens is about 111.7 GB, before the weights.^[3] With Q4_K_M weights that is about 132.5 GB, which needs multiple data-center GPUs. A 128K context needs about 27.9 GB of FP16 cache.

Why does it use more memory than Qwen3.8 27B at the same context?

Qwen3.8 27B keeps a growing cache in only 16 of its 64 layers, while Nanbeige4.3 33B keeps one in all 52 layers, so its cache per token is about 3.25 times larger (212,992 vs 65,536 bytes).^[3] See our Qwen3.8 27B page.

Can a Mac run Nanbeige4.3 33B?

Yes. A Mac with 32 GB of unified memory can run Q5_K_M (25.1 GB) at a moderate context, and a 64 GB Mac can run Q8_0 (35.2 GB) with plenty of room for context.^[4][11]

How we calculated

• Weights are the exact GGUF file sizes on Hugging Face, with multi-part files added together.^[4] BF16 is the two-part BF16 GGUF.
• KV cache is calculated from config.json (52 layers, 8 KV heads, head size 128). We assume an FP16 cache.^[3]
• Device memory is from NVIDIA and Apple spec pages. We compare in decimal GB, so the results lean conservative.
• Not included: engine working buffers, the OS and display. Leave headroom.

Sources

All sources checked on 2026-10-26.

1. Hugging Face, Nanbeige/Nanbeige4.3-33B (parameter count, license, repo date, downloads; checked October 26, 2026). https://huggingface.co/Nanbeige/Nanbeige4.3-33B
2. Hugging Face, Nanbeige4.3-33B model card (benchmarks, 512K context, vLLM and SGLang examples). https://huggingface.co/Nanbeige/Nanbeige4.3-33B/blob/main/README.md
3. Hugging Face, config.json (layers, KV heads, head size, context). https://huggingface.co/Nanbeige/Nanbeige4.3-33B/blob/main/config.json
4. Hugging Face, bartowski/Nanbeige_Nanbeige4.3-33B-GGUF file list. https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF/tree/main
5. NVIDIA, GeForce RTX 3060 family (12 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3060-3060ti/
6. NVIDIA, GeForce RTX 5080 (16 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5080/
7. NVIDIA, GeForce RTX 3090 specs (24 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/
8. NVIDIA, GeForce RTX 4090 specs (24 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/
9. NVIDIA, GeForce RTX 5090 specs (32 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/
10. NVIDIA, RTX 6000 Ada Generation (48 GB). https://www.nvidia.com/en-us/products/workstations/rtx-6000/
11. Apple, MacBook Pro tech specs (unified memory options). https://www.apple.com/macbook-pro/specs/
12. Runpod, GPU pricing (checked October 23, 2026). https://www.runpod.io/pricing
13. Nanbeige on X, Nanbeige4.3-33B announcement (October 25, 2026). https://x.com/nanbeige/status/2114328933181333798
14. @lxe on X, Nanbeige4.3-33B on one RTX 5090 (third-party report, October 26, 2026). https://x.com/lxe/status/2114867943020577065