Kwaipilot_KAT-Coder-V2.5-Dev GGUF size and VRAM requirements
bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF is a very large code-focused language model with 34.66 billion parameters, built on the qwen35moe architecture. It is released under the apache-2.0 license and has been downloaded 396,328 times.
Kwaipilot_KAT-Coder-V2.5-Dev-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 40 transformer layers, a hidden size of 2,048, 16 attention heads. It uses a hybrid attention design: only 10 of its 40 layers keep a full KV cache, while the rest use linear attention with a constant-size state. Long contexts therefore cost far less memory than on a conventional model of this size.
To run bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF locally at a 4,096-token context, its quantized versions need between 9.99 GB (IQ2_XXS, lowest quality) and 65.49 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 Kwaipilot_KAT-Coder-V2.5-Dev-GGUF is about 0.08 GB, rising to roughly 5.0 GB at its full 262,144-token context. Shorter prompts free up memory for a higher-quality quantization.
For most users the best balance is Q5_K_S, needing about 23.38 GB. That means bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF fits entirely in the VRAM of a 10 GB GPU or larger, running fully on the GPU.
Available GGUF quantizations for bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF include IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, Q2_K_L, IQ3_XXS, Q3_K_S, IQ3_XS, Q3_K_M, Q3_K_L, IQ3_M, 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 262,144 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 |
|---|---|---|---|---|---|---|---|
| IQ2_XXS | 2.26 | Very low | 9.11 GB | 0.08 GB | 9.99 GB | 5.5 t/s | Offload |
| IQ2_XS | 2.49 | Very low | 10.06 GB | 0.08 GB | 10.93 GB | 5.0 t/s | Offload |
| IQ2_S | 2.54 | Very low | 10.25 GB | 0.08 GB | 11.13 GB | 4.9 t/s | Offload |
| IQ2_M | 2.79 | Low | 11.24 GB | 0.08 GB | 12.12 GB | 4.4 t/s | Offload |
| Q2_K | 2.91 | Low | 11.75 GB | 0.08 GB | 12.63 GB | 4.3 t/s | Offload |
| Q2_K_L | 3.03 | Low | 12.21 GB | 0.08 GB | 13.09 GB | 4.1 t/s | Offload |
| IQ3_XXS | 3.43 | Fair | 13.85 GB | 0.08 GB | 14.73 GB | 3.6 t/s | Offload |
| Q3_K_S | 3.58 | Fair | 14.45 GB | 0.08 GB | 15.33 GB | 3.5 t/s | Offload |
| IQ3_XS | 3.74 | Fair | 15.1 GB | 0.08 GB | 15.98 GB | 3.3 t/s | Offload |
| Q3_K_M | 3.75 | Fair | 15.11 GB | 0.08 GB | 15.99 GB | 3.3 t/s | Offload |
| Q3_K_L | 3.9 | Fair | 15.73 GB | 0.08 GB | 16.6 GB | 3.2 t/s | Offload |
| IQ3_M | 3.9 | Fair | 15.74 GB | 0.08 GB | 16.62 GB | 3.2 t/s | Offload |
| Q3_K_XL | 4.0 | Fair | 16.14 GB | 0.08 GB | 17.02 GB | 3.1 t/s | Offload |
| IQ4_XS | 4.34 | Good | 17.51 GB | 0.08 GB | 18.39 GB | 2.9 t/s | Offload |
| IQ4_NL | 4.58 | Good | 18.5 GB | 0.08 GB | 19.38 GB | 2.7 t/s | Offload |
| Q4_0 | 4.6 | Good | 18.57 GB | 0.08 GB | 19.45 GB | 2.7 t/s | Offload |
| Q4_K_S | 4.75 | Good | 19.18 GB | 0.08 GB | 20.06 GB | 2.6 t/s | Offload |
| Q4_K_M | 4.94 | Good | 19.92 GB | 0.08 GB | 20.8 GB | 2.5 t/s | Offload |
| Q4_K_L | 5.02 | Very good | 20.27 GB | 0.08 GB | 21.15 GB | 2.5 t/s | Offload |
| Q4_1 | 5.07 | Very good | 20.46 GB | 0.08 GB | 21.34 GB | 2.4 t/s | Offload |
| Q5_K_S | 5.58 | Very good | 22.5 GB | 0.08 GB | 23.38 GB | 2.2 t/s | Offload |
| Q5_K_M | 5.77 | Very good | 23.3 GB | 0.08 GB | 24.18 GB | — | Insufficient |
| Q5_K_L | 5.85 | Very good | 23.59 GB | 0.08 GB | 24.47 GB | — | Insufficient |
| Q6_K | 6.94 | Excellent | 27.99 GB | 0.08 GB | 28.87 GB | — | Insufficient |
| Q6_K_L | 6.99 | Excellent | 28.22 GB | 0.08 GB | 29.1 GB | — | Insufficient |
| Q8_0 | 8.52 | Excellent | 34.38 GB | 0.08 GB | 35.26 GB | — | Insufficient |
| BF16 | 16.01 | Excellent | 64.61 GB | 0.08 GB | 65.49 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/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF?
bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF is a code-focused language model with 34.66 billion parameters, based on the qwen35moe architecture. It is released under the apache-2.0 license and distributed as GGUF files for local inference.
Is bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF a Mixture-of-Experts (MoE) model?
Yes. bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-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/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF handle long context?
bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF uses a hybrid attention design: only 10 of its 40 layers keep a full KV cache, and the rest use linear attention with a constant-size state. Long contexts cost far less memory than on a conventional model of this size.
How much VRAM do you need to run bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF?
You need about 9.99 GB of VRAM to run bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF entirely on the GPU using the IQ2_XXS quantization (at a 4,096-token context). Smaller quantizations lower the requirement at the cost of quality.
Can I run bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF on an 8 GB GPU?
Partially. bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF only fits on an 8 GB GPU by offloading part of it to system RAM (with Q5_K_S), which runs but is slower.
Can I run bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF on a 16 GB GPU?
Yes. With 16 GB of VRAM you can run bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF fully on the GPU using Q3_K_M (about 15.99 GB).
Can I run bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF on a 24 GB GPU?
Yes. With 24 GB of VRAM you can run bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF fully on the GPU using Q5_K_S (about 23.38 GB).
What context length does bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF support?
bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF supports a native context length of up to 262,144 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/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF?
For bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF, a strong default is Q4_K_M, which needs about 20.8 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.