K2 Horizon MoVA 36B-A4B — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
A mixture-of-experts model with Mixture-of-Values attention: 37.4B parameters stored (the card rounds to 36B), about 4B active per token, so it generates quickly for its size. It scores 25 on the Artificial Analysis Intelligence Index, second only to Qwen3.8-27B among models that fit one consumer GPU when checked. Its card reports Terminal-Bench 2.1 58.6 against 44.9 for Qwen3.6-35B-A3B. The official Q4_K_M GGUF is 22.4 GB, and its plain grouped-query cache adds about 6.4 GB at 32K context, so this is a 32 GB-card model; a 24 GB card does not hold it at useful context. Apache 2.0, 512K native context.
K2 Horizon MoVA 36B-A4B needs about 23 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Includes the KV cache at 8K context, so it reads higher than the headline figure.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 14.7 GB | ~198 tok/s | Fits | |
| Q3_K_M 3.41 bpw | 18.4 GB | ~180 tok/s | Fits | |
| Q4_K_M 4.83 bpw | 25 GB | ~23 tok/s | Offload | |
| Q5_K_M 5.67 bpw | 28.9 GB | ~21 tok/s | Offload | |
| Q6_K 6.56 bpw | 33.1 GB | ~20 tok/s | Offload | |
| Q8_0 8.50 bpw | 42.1 GB | ~17 tok/s | Offload | |
| F16 16.00 bpw | 77.2 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. K2 Horizon MoVA 36B-A4B VRAM calculator →
Get K2 Horizon MoVA 36B-A4B running
The cheapest catalogued GPU that runs K2 Horizon MoVA 36B-A4B is the AMD Radeon RX 7900 XTX (24 GB).
or compare on Vast.ai from $0.35/hr (typical low · varies)
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run K2 Horizon MoVA 36B-A4B
No Ollama library tag yet. Run the official GGUF with llama.cpp: the files declare the k2-horizon architecture, which mainline llama.cpp supports; an older Ollama or LM Studio build may not load it. The model card's own serving route is vLLM or SGLang.
Specifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 37.4 Billion (4B active)
- Context window
- 524,288
- Architecture
- Mixture-of-Experts (100 routed + 1 shared) with Mixture-of-Values attention
- Provider
- MBZUAI Institute of Foundation Models
- Licence
- Apache 2.0
- Specified at
- Q4_K_M
- System RAM
- 64 GB
- Record updated
- 2026-10-08
Commercial use permitted. No usage restrictions beyond attribution.
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
| Benchmark | Score | Provenance |
|---|---|---|
| Terminal-Bench 2.1 | 58.6 / 100 % | reported · https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B |
Other K2 Horizon sizes
K2 Horizon MoVA 36B-A4B — frequently asked questions
How much VRAM does K2 Horizon MoVA 36B-A4B need?
About 23 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does K2 Horizon MoVA 36B-A4B run on an RTX 4090 (24 GB)?
Yes. K2 Horizon MoVA 36B-A4B needs about 23 GB at Q4_K_M, inside a 24 GB card, at an estimated 23 tokens/sec.
How do I run K2 Horizon MoVA 36B-A4B locally?
No Ollama library tag yet. Run the official GGUF with llama.cpp: the files declare the k2-horizon architecture, which mainline llama.cpp supports; an older Ollama or LM Studio build may not load it. The model card's own serving route is vLLM or SGLang. Running the published tag would send your prompts to a hosted GPU rather than your own machine.
What other sizes does K2 Horizon come in?
K2 Horizon 3.7B (4 GB), K2 Horizon 7B (6 GB), K2 Horizon MoVA 36B-A4B (23 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.