llmrun
sahilchachraQwen3_5MoeForConditionalGeneration

KAT Coder V2.5 Dev NVFP4A16 — Hardware Requirements & GPU Compatibility

ChatCodeFunctions

KAT Coder V2.5 Dev NVFP4A16 is a 20.4B-parameter open language model from sahilchachra. It supports a context window of up to 262,144 tokens. At Q4_K_M it needs about 12.62 GB of VRAM — see which GPUs and Macs can run it below.

20.4B params · dense262K contextApache 2.079.8K downloads2 likes
A

Runs well on a 24 GB GPU

Runs from12 GB Q3_K_M · 128KBest on24 GB Q4_K_M · ~52 tok/s · 256KMacs from16 GB Q3_K_M · ~14 tok/s

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run KAT Coder V2.5 Dev NVFP4A16

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Recommended

vLLM
one command · needs vLLM installed
$vllm serve sahilchachra/KAT-Coder-V2.5-Dev-NVFP4A16 --max-model-len 16384

vLLM repo with its own quantisation

Download: Weights on Hugging Face

install, OS, other quants

Ways to runOther ways

each opens step-by-step

Run in cloud

Runs on RTX 3090 Ti (24 GB) (11 GB headroom) · Q4_K_M

Generation speed
~52 tok/s
generation speed
Cost per 1M output tokens
$1.30
per 1M output tokens

How Much VRAM Does KAT Coder V2.5 Dev NVFP4A16 Need?

Select a quantization level; the listed hardware adheres to it.

QuantizationVRAM+32KOn RTX 4090 (reference)
9.1 GB9.3 GB2-bit · K-quant — 2-bit quantization with K-quant improvementsFits: 256K context · ~72 tok/s
10.3 GB10.6 GB3-bit medium — 3-bit medium quantizationFits: 256K context · ~63 tok/s
12.6 GB12.9 GB4-bit medium — 4-bit medium quantization — most popular sweet spotFits: 256K context · ~52 tok/s
14.9 GB15.2 GB5-bit medium — 5-bit medium quantization — good quality/size tradeoffFits: 256K context · ~44 tok/s
17.2 GB17.5 GB6-bit, very good — 6-bit quantization, very good qualityFits: 256K context · ~38 tok/s
20.8 GB21 GB8-bit, near-lossless — 8-bit quantization, near-losslessFits: 256K context · ~32 tok/s
41.2 GB41.4 GB16-bit · bfloat — Brain floating point 16 — preferred for trainingDoes not fit: needs 41.3 GB
est.calculated, no published file yet

Selected Q4_K_M

4-bit medium quantization

12.7GBat 16K context

Weights
12.2 GB
KV cache
0.2 GB
Overhead
0.3 GB

Fits RTX 4090 (reference) · 11.3 GB free · ~52 tok/s

≥50 tok/s needs ≥970 GB/s · RTX 5090 ~92, M5 Ultra ~68

Which Hardware Can Run KAT Coder V2.5 Dev NVFP4A16?

Q4_K_M · 12.6 GB

KAT Coder V2.5 Dev NVFP4A16 (Q4_K_M) requires 12.6 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 17+ GB is recommended. Using the full 262K context window can add up to 2.6 GB, bringing total usage to 15.2 GB. 30 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 5080. 140 devices with unified memory can run KAT Coder V2.5 Dev NVFP4A16, including Mac Studio M5 Ultra (256 GB), Mac Studio M5 Ultra (512 GB), MacBook Pro 14" M3 Pro (18 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M
8 GB · RTX 4060Doesn't fit — smallest quant needs 9.1 GB8 GB tier →
12 GB · RTX 3060 12GBQ3_K_M · 10.4 GB · ~23 tok/s · 128K12 GB tier →
16 GB · RTX 5060 Ti 16GBQ4_K_M · 12.7 GB · ~23 tok/s · 256K16 GB tier →
24 GB · RTX 4090referenceQ4_K_M · 12.7 GB · ~52 tok/s · 256K24 GB tier →
Mac 32 GB · Mac Mini M4Q4_K_M · 12.7 GB · ~7 tok/s · 256KMacs →
Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 12.7 GB · ~30 tok/s · 256KMacs →

Alternatives to KAT Coder V2.5 Dev NVFP4A16

Similar models

3
ModelScoreVRAMtok/sWhy it is similar
InternVL3 5 GPT OSS 20B A4B Preview HFOpenGVLab—13.6 GB~180same size
Internlm 20BInternLM—45.3 GB—needs > 24 GBsame size
ERNIE 4.5 21B A3B ThinkingBaidu—14.3 GB~213same size

Specifications

Publisher
sahilchachra
Parameters
20.4B
Architecture
Qwen3_5MoeForConditionalGeneration
Context Length
262,144 tokens
Vocabulary Size
248,320
Release Date
2026-07-24
License
Apache 2.0

Frequently Asked Questions

How much VRAM does KAT Coder V2.5 Dev NVFP4A16 need?

KAT Coder V2.5 Dev NVFP4A16 requires 12.6 GB of VRAM at Q4_K_M, or 41.2 GB at BF16. Full 262K context adds up to 2.6 GB (15.2 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 20.4B × 4.8 bits ÷ 8 = 12.2 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

12.6 GB
15.2 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run KAT Coder V2.5 Dev NVFP4A16?

Yes, at Q8_0 (20.8 GB) or lower. Higher quantizations like BF16 (41.2 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for KAT Coder V2.5 Dev NVFP4A16?

For KAT Coder V2.5 Dev NVFP4A16, Q4_K_M (12.6 GB) offers the best balance of quality and VRAM usage. Q5_K_M (14.9 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 9.1 GB.

VRAM requirement by quantization

Q2_K
9.1 GB
Q4_K_M ★
12.6 GB
Q5_K_M
14.9 GB
Q6_K
17.2 GB
Q8_0
20.8 GB
BF16
41.2 GB

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run KAT Coder V2.5 Dev NVFP4A16 on a Mac?

Yes — Mac Mini M2 Pro (16 GB) and 145 other Macs can run KAT Coder V2.5 Dev NVFP4A16, the smallest at Q3_K_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 12.6 GB of usable unified memory (RAM minus macOS overhead), from MacBook Pro 14" M3 Pro (18 GB).

Can I run KAT Coder V2.5 Dev NVFP4A16 locally?

Yes — KAT Coder V2.5 Dev NVFP4A16 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is KAT Coder V2.5 Dev NVFP4A16?

At Q4_K_M, KAT Coder V2.5 Dev NVFP4A16 can reach ~380 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~52 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = (bandwidth GB/s ÷ model GB) × efficiency

Example: NVIDIA B200 → 8000 ÷ 12.62 × 0.65 = ~412 tok/s

Estimated speed at Q4_K_M (12.6 GB)

~412 tok/s
~52 tok/s
~412 tok/s
~380 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of KAT Coder V2.5 Dev NVFP4A16?

At Q4_K_M, the download is about 12.24 GB. The full-precision BF16 version is 40.79 GB. The smallest option (Q2_K) is 8.67 GB.

Which GPUs can run KAT Coder V2.5 Dev NVFP4A16?

30 consumer GPUs can run KAT Coder V2.5 Dev NVFP4A16 at Q4_K_M. Top options include AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090, AMD Radeon RX 6800. 8 GPUs have plenty of headroom for comfortable inference.

Which devices can run KAT Coder V2.5 Dev NVFP4A16?

140 devices with unified memory can run KAT Coder V2.5 Dev NVFP4A16 at Q4_K_M (12.6 GB), including AMD Ryzen 7 8845HS (Hawk Point) Laptop, AMD Ryzen AI 7 350 (Krackan Point) Laptop, AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10. Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.