A.X K2 — Hardware Requirements & GPU Compatibility
ChatFunctions
A.X K2 is a 691.7B-parameter open language model from skt. It supports a context window of up to 262,144 tokens. At Q4_K_M it needs about 418.90 GB of VRAM — see which GPUs and Macs can run it below.
A.X K2 runs on a 512 GB+ Mac or a multi-GPU server
Best local pick → Mac Studio (M3 Ultra, 512GB) · Q4_K_M · 444 GB · ~20 tok/s
Macs from 512 GB (Q4_K_M · ~20 tok/s)
Apache 2.0
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run A.X K2
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 serve skt/A.X-K2 --max-model-len 16384 --tensor-parallel-size 8Size the nodes from the table above
Download: A.X-K2-Q4_K_M.gguf415.01 GBskt
Ways to runOther ways
each opens step-by-stepRun in cloud
Rent 7× H100 SXM · Q4_K_M
- Estimated hourly price
- ~$18.83/hr
- est., whole server
- GPUs
- 7 GPUs
- 560 GB of VRAM
$5+ bonus for new users
How Much VRAM Does A.X K2 Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | Runs on | |
|---|---|---|---|---|
| 297.9 GB | 351.6 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Runs on 5× H100 SXM 80 GB | |
| 341.1 GB | 394.8 GB | 3-bit medium — 3-bit medium quantization | Runs on 6× H100 SXM 80 GB | |
| 375.7 GB | 429.4 GB | 4-bit · i-quant compact — Importance-weighted 4-bit, compact | Runs on 6× H100 SXM 80 GB | |
| 418.9 GB | 472.6 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Runs on 7× H100 SXM 80 GB | |
| 496.7 GB | 550.4 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Runs on 8× H100 SXM 80 GB | |
| 574.5 GB | 628.3 GB | 6-bit, very good — 6-bit quantization, very good quality | Runs on 5× H200 SXM 141 GB | |
| 695.6 GB | 749.3 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Runs on 6× H200 SXM 141 GB | |
| 1387.3 GB | 1441 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Runs on 3 nodes of 8× H100 SXM |
Selected Q4_K_M
4-bit medium quantization444.0GBat 16K context
- Weights
- 415.0 GB
- KV cache
- 28.7 GB
- Overhead
- 0.3 GB
Server-class: ~444 GB at 16K context · about 7× H100 SXM 80 GB
Which Hardware Can Run A.X K2?
Q4_K_M · 418.9 GBA.X K2 (Q4_K_M) requires 418.9 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 545+ GB is recommended. Using the full 262K context window can add up to 454.9 GB, bringing total usage to 873.8 GB. No consumer GPU has enough memory. 1 device with unified memory can run A.X K2, including Mac Studio (M3 Ultra, 512GB).
Needs server-class hardware
A.X K2 runs on a 512 GB+ Mac or a multi-GPU server · Q4_K_M needs ~444 GB
| Hardware | Quant | Memory | Speed | Context | Hardware page | |
|---|---|---|---|---|---|---|
| Mac 512 GB · Mac Studio (M3 Ultra, 512GB) | Q4_K_M | 444 GB | ~20 tok/s | 32K ctx | Mac Studio → | Q4_K_M · 444 GB · ~20 tok/s · 32K ctx |
| 7× H100 SXM 80 GB · 560 GB | Q4_K_M | 444 GB | — | 32K ctx | H100 SXM → | Q4_K_M · 444 GB · 32K ctx |
| 4× H200 SXM 141 GB · 564 GB | Q4_K_M | 444 GB | — | 32K ctx | H200 SXM → | Q4_K_M · 444 GB · 32K ctx |
| 3× B200 192 GB · 576 GB | Q4_K_M | 444 GB | — | 32K ctx | B200 → | Q4_K_M · 444 GB · 32K ctx |
Also: Mac Studio (M3 Ultra, 512GB) · NVIDIA DGX A100 640GB · NVIDIA DGX H100
By VRAM tier · typical GPUs and Macs at Q4_K_M
Consumer GPUs (up to 32 GB) and Macs up to 256 GB: too small for this model (needs ≥ 322.9 GB)8 GB tier →12 GB tier →16 GB tier →24 GB tier →Macs →
Similar models
3| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| DeepSeek V3.2DeepSeek | 110.4 | 412.7 GB | —needs > 24 GB | same size |
| DeepSeek R1 0528DeepSeek | 103.7 | 412.2 GB | —needs > 24 GB | same size |
| Mistral Large 3 675B Instruct 2512Mistral AI | 89.6 | 454.4 GB | —needs > 24 GB | same size |
Specifications
- Publisher
- skt
- Parameters
- 691.7B
- Active parameters
- 36.4B per token
- Architecture
- AXK2ForCausalLM
- Context Length
- 262,144 tokens
- Vocabulary Size
- 163,840
- Release Date
- 2026-07-28
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does A.X K2 need?
A.X K2 requires 418.9 GB of VRAM at Q4_K_M, or 1387.3 GB at BF16. Full 262K context adds up to 454.9 GB (873.8 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 691.7B × 4.8 bits ÷ 8 = 415 GB
KV Cache + Overhead ≈ 3.9 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 ≈ 458.8 GB (at full 262K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 5090 run A.X K2?
No — A.X K2 requires at least 297.9 GB at Q2_K, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
What's the best quantization for A.X K2?
For A.X K2, Q4_K_M (418.9 GB) offers the best balance of quality and VRAM usage. Q5_K_M (496.7 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 297.9 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run A.X K2 on a Mac?
Yes — Mac Studio (M3 Ultra, 512GB) and above can run A.X K2. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 418.9 GB of usable unified memory (RAM minus macOS overhead).
Can I run A.X K2 locally?
Yes — A.X K2 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
What's the download size of A.X K2?
At Q4_K_M, the download is about 415.01 GB. The full-precision BF16 version is 1383.38 GB. The smallest option (Q2_K) is 293.97 GB.
Which GPUs can run A.X K2?
No single consumer GPU has enough VRAM to run A.X K2 at Q4_K_M (418.9 GB). Multi-GPU or professional hardware is required.
Which devices can run A.X K2?
1 device with unified memory can run A.X K2 at Q4_K_M (418.9 GB), including Mac Studio (M3 Ultra, 512GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.