moonshotai_Kimi-Linear-48B-A3B-Instruct GGUF size and VRAM requirements
bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF is a very large instruction-tuned chat model with 49.12 billion parameters, built on the kimi-linear architecture. It is released under the mit license and has been downloaded 10,931 times.
moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF is a Mixture-of-Experts model with 256 experts, of which 8 are active on each token. Routing only 8 of 256 experts makes it noticeably faster than a dense model of the same size — but every expert still has to be held in memory, so the numbers below are set by the full parameter count, not the active one. It is built from 27 transformer layers, a hidden size of 2,304, 32 attention heads. moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF uses Multi-head Latent Attention, caching a single compressed latent per layer instead of full keys and values. That keeps the KV cache small even at long context — the memory figures on this page already reflect it.
To run bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF locally at a 4,096-token context, its quantized versions need between 10.6 GB (IQ1_S, lowest quality) and 92.37 GB (BF16, highest quality) of memory, weights plus KV cache and a system margin included. Context length drives that memory directly: at 4,096 tokens the KV cache for moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF is about 0.03 GB, rising to roughly 7.88 GB at its full 1,048,576-token context. Shorter prompts free up memory for a higher-quality quantization.
For most users the best balance is Q3_K_XL, needing about 23.13 GB. That means bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF fits entirely in the VRAM of a 12 GB GPU or larger, running fully on the GPU.
Available GGUF quantizations for bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF include IQ1_S, IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, Q2_K_L, IQ3_XXS, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, Q3_K_XL, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_K_L, Q4_1, Q5_K_S, Q5_K_M, Q5_K_L, Q6_K, Q6_K_L, Q8_0, BF16. The model supports a native context length of up to 1,048,576 tokens; a longer context grows the KV cache and the memory needed.
GGUF file size and memory by quantization
Compare real GGUF weight sizes, estimated KV cache and total memory for Q4, Q5, Q8 and every quantization published in this repository.
| Quant. | Bits | Quality | Weights | KV | Total | Speed~ | Verdict |
|---|---|---|---|---|---|---|---|
| IQ1_S | 1.71 | Very low | 9.77 GB | 0.03 GB | 10.6 GB | 5.1 t/s | Offload |
| IQ1_M | 1.78 | Very low | 10.17 GB | 0.03 GB | 11.0 GB | 4.9 t/s | Offload |
| IQ2_XXS | 1.98 | Very low | 11.3 GB | 0.03 GB | 12.13 GB | 4.4 t/s | Offload |
| IQ2_XS | 2.26 | Very low | 12.94 GB | 0.03 GB | 13.77 GB | 3.9 t/s | Offload |
| IQ2_S | 2.27 | Very low | 13.01 GB | 0.03 GB | 13.84 GB | 3.8 t/s | Offload |
| IQ2_M | 2.57 | Very low | 14.71 GB | 0.03 GB | 15.54 GB | 3.4 t/s | Offload |
| Q2_K | 2.86 | Low | 16.33 GB | 0.03 GB | 17.16 GB | 3.1 t/s | Offload |
| Q2_K_L | 2.92 | Low | 16.67 GB | 0.03 GB | 17.5 GB | 3.0 t/s | Offload |
| IQ3_XXS | 3.2 | Low | 18.3 GB | 0.03 GB | 19.13 GB | 2.7 t/s | Offload |
| IQ3_XS | 3.33 | Fair | 19.04 GB | 0.03 GB | 19.87 GB | 2.6 t/s | Offload |
| Q3_K_S | 3.52 | Fair | 20.12 GB | 0.03 GB | 20.95 GB | 2.5 t/s | Offload |
| IQ3_M | 3.69 | Fair | 21.1 GB | 0.03 GB | 21.93 GB | 2.4 t/s | Offload |
| Q3_K_M | 3.69 | Fair | 21.12 GB | 0.03 GB | 21.95 GB | 2.4 t/s | Offload |
| Q3_K_L | 3.85 | Fair | 22.0 GB | 0.03 GB | 22.83 GB | 2.3 t/s | Offload |
| Q3_K_XL | 3.9 | Fair | 22.3 GB | 0.03 GB | 23.13 GB | 2.2 t/s | Offload |
| IQ4_XS | 4.31 | Good | 24.65 GB | 0.03 GB | 25.48 GB | — | Insufficient |
| IQ4_NL | 4.56 | Good | 26.05 GB | 0.03 GB | 26.88 GB | — | Insufficient |
| Q4_0 | 4.63 | Good | 26.49 GB | 0.03 GB | 27.32 GB | — | Insufficient |
| Q4_K_S | 4.72 | Good | 26.99 GB | 0.03 GB | 27.82 GB | — | Insufficient |
| Q4_K_M | 4.9 | Good | 28.0 GB | 0.03 GB | 28.83 GB | — | Insufficient |
| Q4_K_L | 4.94 | Good | 28.26 GB | 0.03 GB | 29.09 GB | — | Insufficient |
| Q4_1 | 5.04 | Very good | 28.85 GB | 0.03 GB | 29.68 GB | — | Insufficient |
| Q5_K_S | 5.54 | Very good | 31.68 GB | 0.03 GB | 32.51 GB | — | Insufficient |
| Q5_K_M | 5.72 | Very good | 32.69 GB | 0.03 GB | 33.52 GB | — | Insufficient |
| Q5_K_L | 5.75 | Very good | 32.9 GB | 0.03 GB | 33.73 GB | — | Insufficient |
| Q6_K | 6.59 | Excellent | 37.68 GB | 0.03 GB | 38.51 GB | — | Insufficient |
| Q6_K_L | 6.62 | Excellent | 37.85 GB | 0.03 GB | 38.68 GB | — | Insufficient |
| Q8_0 | 8.51 | Excellent | 48.66 GB | 0.03 GB | 49.49 GB | — | Insufficient |
| BF16 | 16.01 | Excellent | 91.54 GB | 0.03 GB | 92.37 GB | — | Insufficient |
KV cache computed from the model's exact architecture. Speed is a rough estimate bounded by memory bandwidth.
Frequently asked questions
What kind of model is bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF?
bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF is an instruction-tuned chat model with 49.12 billion parameters, based on the kimi-linear architecture. It is released under the mit license and distributed as GGUF files for local inference.
Is bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF a Mixture-of-Experts (MoE) model?
Yes. bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF is a Mixture-of-Experts model with 256 experts, of which 8 are activated per token. That makes it faster than a dense model of the same size, but all 256 experts must be loaded into memory, so the VRAM/RAM it needs is driven by the total parameter count, not the active one.
How does bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF handle long context?
bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF uses Multi-head Latent Attention, which caches one compressed latent per layer instead of full keys and values. Its KV cache therefore grows much more slowly than a standard model of the same size, so long contexts add relatively little memory.
How much VRAM do you need to run bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF?
You need about 11.0 GB of VRAM to run bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF entirely on the GPU using the IQ1_M quantization (at a 4,096-token context). Smaller quantizations lower the requirement at the cost of quality.
Can I run bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF on an 8 GB GPU?
Partially. bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF only fits on an 8 GB GPU by offloading part of it to system RAM (with Q3_K_XL), which runs but is slower.
Can I run bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF on a 16 GB GPU?
Yes. With 16 GB of VRAM you can run bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF fully on the GPU using IQ2_M (about 15.54 GB).
Can I run bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF on a 24 GB GPU?
Yes. With 24 GB of VRAM you can run bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF fully on the GPU using Q3_K_XL (about 23.13 GB).
What context length does bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF support?
bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF supports a native context length of up to 1,048,576 tokens. A longer context grows the KV cache, so it increases the memory needed to run the model.
What is the best quantization for bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF?
For bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF, a strong default is Q4_K_M, which needs about 28.83 GB and keeps most of the quality while roughly halving the memory versus 8-bit. With VRAM to spare, Q5_K_M or Q6_K add a little more quality; if you are tight on memory, a smaller quantization still runs. Pick the highest quantization that fits your VRAM.