KAT Coder V2.5 Dev — Hardware Requirements & GPU Compatibility
ChatCodeFunctions
KAT Coder V2.5 Dev is a 34.7B-parameter open language model from Kwaipilot. It supports a context window of up to 262,144 tokens. At Q4_K_M it needs about 21.18 GB of VRAM — see which GPUs and Macs can run it below.
Runs great on a 24 GB GPU
Runs from 12 GB (IQ2_S · 64K)Best on 24 GB (Q4_K_M · ~184 tok/s · 256K)Macs from 16 GB (IQ2_M · ~60 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run KAT Coder V2.5 Dev
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
ollama run hf.co/bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 2.7 GB free · 16K context
Download: Kwaipilot_KAT-Coder-V2.5-Dev-Q4_K_M.gguf21.39 GBbartowski
Other ways
each opens step-by-stepRun in cloud
Fits on RTX PRO 4500 (32 GB) (10 GB headroom) · Q4_K_M
- Generation speed
- ~172 tok/s
- generation speed
- Cost per 1M output tokens
- $0.55
- per 1M output tokens
How Much VRAM Does KAT Coder V2.5 Dev Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 15.1 GB | 15.4 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 256K context · ~222 tok/s | |
| 15.6 GB | 15.8 GB | 3-bit small — 3-bit small quantization | Fits: 256K context · ~219 tok/s | |
| 17.3 GB | 17.5 GB | 3-bit medium — 3-bit medium quantization | Fits: 256K context · ~207 tok/s | |
| 17.7 GB | 18 GB | 4-bit · legacy — 4-bit legacy quantization | Fits: 256K context · ~204 tok/s | |
| 21.2 GB | 21.4 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 256K context · ~184 tok/s | |
| 25.1 GB | 25.3 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Does not fit: needs 25.2 GB | |
| 29.0 GB | 29.2 GB | 6-bit, very good — 6-bit quantization, very good quality | Does not fit: needs 29.1 GB | |
| 35.0 GB | 35.3 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 35.1 GB |
Selected Q4_K_M
4-bit medium quantization21.3GBat 16K context
- Weights
- 20.8 GB
- KV cache
- 0.2 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 2.7 GB free · ~184 tok/s
≥50 tok/s needs ≥182 GB/s · RTX 5090 ~249, RTX 3090 Ti ~184
Which Hardware Can Run KAT Coder V2.5 Dev?
Q4_K_M · 21.2 GBKAT Coder V2.5 Dev (Q4_K_M) requires 21.2 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 28+ GB is recommended. Using the full 262K context window can add up to 2.6 GB, bringing total usage to 23.8 GB. 7 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 100 devices with unified memory can run KAT Coder V2.5 Dev, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M1 Max (32 GB).
20 GB
Q4_K_M · best fit Q4_1 (20.0 GB)
1 GPU
RX 7900 XT~83Simulate output on RX 7900 XT32–40 GB
fits · 10.7 GB free · up to 256K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~249Simulate output on RTX 5090M5 Max36 GB~69Simulate output on Mac Studio M5 Max (36 GB)M532 GB~36Simulate output on MacBook Air 13" M5 (32 GB)Orin~55Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~36Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Qwen3.6 35B A3BAlibaba | 106.8 | 22 GB | ~151 | same size |
| MiniCPM V 4.7 35B A3BOpenBMB | — | 22.1 GB | ~169 | same size |
| Ornith 1.5 35B A3Bornith-ai | — | 22 GB | ~151 | same size |
| Agents A1InternScience | — | 21.5 GB | ~171 | same size |
| Qwen AgentWorld 35B A3BAlibaba | — | 21.3 GB | ~184 | same size |
| Intern S2 PreviewInternLM | — | 73.2 GB | —needs > 24 GB | same size |
Fine-tunes & variants(2)
Specifications
- Publisher
- Kwaipilot
- Parameters
- 34.7B
- Active parameters
- 3.5B per token
- Architecture
- Qwen3_5MoeForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 248,320
- Release Date
- 2026-07-23
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does KAT Coder V2.5 Dev need?
KAT Coder V2.5 Dev requires 21.2 GB of VRAM at Q4_K_M, or 69.7 GB at BF16. Full 262K context adds up to 2.6 GB (23.8 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 34.7B × 4.8 bits ÷ 8 = 20.8 GB
KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 3 GB (at full 262K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run KAT Coder V2.5 Dev?
Yes, at Q4_K_L (21.6 GB) or lower. Higher quantizations like Q5_K_S (24.2 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for KAT Coder V2.5 Dev?
For KAT Coder V2.5 Dev, Q4_K_M (21.2 GB) offers the best balance of quality and VRAM usage. Q4_K_L (21.6 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 9.9 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run KAT Coder V2.5 Dev on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run KAT Coder V2.5 Dev, the smallest at IQ2_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 21.2 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio M1 Max (32 GB).
Can I run KAT Coder V2.5 Dev locally?
Yes — KAT Coder V2.5 Dev can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is KAT Coder V2.5 Dev?
At Q4_K_M, KAT Coder V2.5 Dev can reach ~119 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~184 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 2.1 GB active ÷ (8000 × 0.65) = 0.406 ms, plus 40 layers × 0.055 ms = 2.200 ms, so 1000 ÷ 2.606 ms = ~384 tok/s
Estimated speed at Q4_K_M (21.2 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of KAT Coder V2.5 Dev?
At Q4_K_M, the download is about 20.80 GB. The full-precision BF16 version is 69.32 GB. The smallest option (IQ2_XXS) is 9.53 GB.
Which GPUs can run KAT Coder V2.5 Dev?
7 consumer GPUs can run KAT Coder V2.5 Dev at Q4_K_M. Top options include NVIDIA GeForce RTX 5090, AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090. 1 GPU has plenty of headroom for comfortable inference.
Which devices can run KAT Coder V2.5 Dev?
100 devices with unified memory can run KAT Coder V2.5 Dev at Q4_K_M (21.2 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.