IFMK2HorizonForCausalLM

K2 Horizon MoVA 36B A4B — Hardware Requirements & GPU Compatibility

ChatFunctions

K2 Horizon MoVA 36B A4B is a 37.4B-parameter open language model from IFM. It supports a context window of up to 524,288 tokens. At Q4_K_M it needs about 23.02 GB of VRAM — see which GPUs and Macs can run it below.

37.4B params · MoE4B active524K contextApache 2.0Sep 202615.4K downloads418 likes55.7K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

B

Decent on a 24 GB GPU

Runs from 16 GB (IQ2_M · 16K)Best on 24 GB (Q4_K_M · ~156 tok/s · 8K)Macs from 18 GB (IQ2_S · ~45 tok/s)

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run K2 Horizon MoVA 36B A4B

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Recommended

Ollama

one command · needs Ollama installed

$ollama run hf.co/IFM/K2-Horizon-MoVA-36B-A4B-GGUF:Q4_K_M

Q4_K_M needs offload on the RTX 4090 · 0.8 GB short · 16K context

Download: K2-Horizon-MoVA-36B-A4B-Q4_K_M.gguf22.37 GBIFM

install, OS, other quants

Other ways

each opens step-by-step

Run in cloud

Fits on RTX PRO 4500 (32 GB) (7 GB headroom) · Q4_K_M

Generation speed: ~146 tok/s; generation speed
Cost per 1M output tokens: $0.65; per 1M output tokens

Deployon RunPodBest optionfrom $0.34/hr

Compare GPUs →

or

How Much VRAM Does K2 Horizon MoVA 36B A4B Need?

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

Quantization	VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
	10.8 GB	14.6 GB	26.7 GB	10.1 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 64K context · ~227 tok/s
	11.8 GB	15.6 GB	27.6 GB	11.2 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 64K context · ~219 tok/s
	12.3 GB	16 GB	28.1 GB	11.4 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 64K context · ~216 tok/s
	13.2 GB	17 GB	29 GB	12.5 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 64K context · ~209 tok/s
	15.1 GB	18.8 GB	30.9 GB	14.6 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 64K context · ~196 tok/s
	15.5 GB	19.3 GB	31.4 GB	11.1 GB	2-bit · small K-quant — 2-bit small K-quant	Fits: 64K context · ~193 tok/s
	16 GB	19.8 GB	31.9 GB	15.7 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 64K context · ~191 tok/s
	16.5 GB	20.2 GB	32.3 GB	13.6 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 32K context · ~188 tok/s
	16.5 GB	20.2 GB	32.3 GB	16.4 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 32K context · ~188 tok/s
	16.9 GB	20.7 GB	32.8 GB	16.4 GB	3-bit small — 3-bit small quantization	Fits: 32K context · ~185 tok/s
	17.4 GB	21.2 GB	33.3 GB	16.5 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 32K context · ~183 tok/s
	18.8 GB	22.6 GB	34.7 GB	17.7 GB	3-bit medium — 3-bit medium quantization	Fits: 32K context · ~175 tok/s
	19.3 GB	23 GB	35.1 GB	21.3 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 32K context · ~173 tok/s
	19.7 GB	23.5 GB	35.6 GB	18.8 GB	3-bit large — 3-bit large quantization	Fits: 32K context · ~171 tok/s
	20.7 GB	24.5 GB	36.5 GB	20.1 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 16K context · ~166 tok/s
	21.6 GB	25.4 GB	37.5 GB	21.4 GB	4-bit small — 4-bit small quantization	Fits: 16K context · ~162 tok/s
	21.6 GB	25.4 GB	37.5 GB	23.6 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Fits: 16K context · ~162 tok/s
	21.6 GB	25.4 GB	37.5 GB	21.3 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Fits: 16K context · ~162 tok/s
	23.0 GB	26.8 GB	38.9 GB	22.4 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 8K context · ~156 tok/s
	23.9 GB	27.7 GB	39.8 GB	25.9 GB	5-bit · legacy — 5-bit legacy quantization	Fits: 2K context · ~153 tok/s
	26.3 GB	30.1 GB	42.1 GB	25.9 GB	5-bit small — 5-bit small quantization	Does not fit: needs 28.1 GB
	26.3 GB	30.1 GB	42.1 GB	28.2 GB	5-bit · legacy+ — 5-bit legacy quantization with offset	Does not fit: needs 28.1 GB
	27.2 GB	31 GB	43.1 GB	26.4 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Does not fit: needs 29 GB
	31.4 GB	35.2 GB	47.3 GB	30.8 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 33.2 GB
	38 GB	41.8 GB	53.9 GB	39.8 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 39.8 GB
	75.4 GB	79.2 GB	91.3 GB	74.9 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 77.2 GB

Selected Q4_K_M

4-bit medium quantization

24.8GBat 16K context

Weights: 22.5 GB
KV cache: 2.0 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · fits up to 8K context · ~156 tok/s

≥50 tok/s needs ≥218 GB/s · RTX 5090 ~211, M3 Ultra ~72

Or rent RTX PRO 4500 from $0.34/hr.

Which Hardware Can Run K2 Horizon MoVA 36B A4B?

Q4_K_M · 23.0 GB

K2 Horizon MoVA 36B A4B (Q4_K_M) requires 23.0 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 30+ GB is recommended. Using the full 524K context window can add up to 64.2 GB, bringing total usage to 87.2 GB. 7 GPUs can run it, including NVIDIA GeForce RTX 5090. 100 devices with unified memory can run K2 Horizon MoVA 36B A4B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M5 Max (36 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8–12 GB · RTX 4060, RTX 3060 12GBDoesn't fit — smallest quant needs 12.6 GB8 GB tier →12 GB tier →

16 GB · RTX 5060 Ti 16GBIQ2_M · 15 GB · ~134 tok/s · 16K15~13416K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 24.8 GB · ~156 tok/s · 8Ktight24.8~1568K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 24.8 GB · ~26 tok/s · 32K24.8~2632KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 24.8 GB · ~62 tok/s · 256K24.8~62256KMacs →

Alternatives to K2 Horizon MoVA 36B A4B

Similar models

6tok/s on RTX 4090

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
Intern S2 PreviewInternLM	—	73.2 GB	—needs > 24 GB	same size
Qwen3.5 35B A3B BaseAlibaba	—	22 GB	~151	same size

Fine-tunes & variants(1)community builds of K2 Horizon MoVA 36B A4B

K2 Horizon MoVA 36B A4B Uncensored

Specifications

Publisher: IFM
Parameters: 37.4B
Active parameters: 4B per token
Architecture: K2HorizonForCausalLM
Context Length: 524,288 tokens
Vocabulary Size: 250,624
Release Date: 2026-09-01
License: Apache 2.0

Resources

Hugging Face model card (opens in a new tab)

Frequently Asked Questions

How much VRAM does K2 Horizon MoVA 36B A4B need?

K2 Horizon MoVA 36B A4B requires 23.0 GB of VRAM at Q4_K_M, or 75.4 GB at BF16. Full 524K context adds up to 64.2 GB (87.2 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 37.4B × 4.8 bits ÷ 8 = 22.5 GB

KV Cache + Overhead ≈ 0.5 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 ≈ 64.7 GB (at full 524K context)

VRAM usage by quantization

Q4_K_M

23.0 GB

Q4_K_M + full context

87.2 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run K2 Horizon MoVA 36B A4B?

Yes, at Q5_0 (23.9 GB) or lower. Higher quantizations like Q5_K_S (26.3 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for K2 Horizon MoVA 36B A4B?

For K2 Horizon MoVA 36B A4B, Q4_K_M (23.0 GB) offers the best balance of quality and VRAM usage. Q5_0 (23.9 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 10.8 GB.

VRAM requirement by quantization

IQ2_XXS

10.8 GB~53%

IQ3_XS

16.0 GB~73%

Q3_K_L

19.7 GB~86%

Q4_K_M ★

23.0 GB~89%

Q5_0

23.9 GB~90%

BF16

75.4 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run K2 Horizon MoVA 36B A4B on a Mac?

Yes — MacBook Pro 14" M3 Pro (18 GB) and 108 other Macs can run K2 Horizon MoVA 36B A4B, the smallest at IQ2_S. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 23.0 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio M1 Max (32 GB).

Can I run K2 Horizon MoVA 36B A4B locally?

Yes — K2 Horizon MoVA 36B A4B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is K2 Horizon MoVA 36B A4B?

At Q4_K_M, K2 Horizon MoVA 36B A4B can reach ~99 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~156 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 → 2.5 GB active ÷ (8000 × 0.65) = 0.473 ms, plus 48 layers × 0.055 ms = 2.640 ms, so 1000 ÷ 3.113 ms = ~321 tok/s

Estimated speed at Q4_K_M (23.0 GB)

NVIDIA B200

~321 tok/s

NVIDIA GeForce RTX 4090

~156 tok/s

NVIDIA B300

~321 tok/s

NVIDIA GH200 Grace Hopper Superchip

~293 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 K2 Horizon MoVA 36B A4B?

At Q4_K_M, the download is about 22.47 GB. The full-precision BF16 version is 74.89 GB. The smallest option (IQ2_XXS) is 10.30 GB.

Which GPUs can run K2 Horizon MoVA 36B A4B?

7 consumer GPUs can run K2 Horizon MoVA 36B A4B at Q4_K_M. Top options include AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090.

Which devices can run K2 Horizon MoVA 36B A4B?

100 devices with unified memory can run K2 Horizon MoVA 36B A4B at Q4_K_M (23.0 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.

Run K2 Horizon MoVA 36B A4B

24.8 GB VRAM at 16K context · needs offload on the RTX 4090 (0.8 GB short)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/IFM/K2-Horizon-MoVA-36B-A4B-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf IFM/K2-Horizon-MoVA-36B-A4B-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/IFM/K2-Horizon-MoVA-36B-A4B-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf IFM/K2-Horizon-MoVA-36B-A4B-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve IFM/K2-Horizon-MoVA-36B-A4B --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~82 GB of GPU memory.

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$IFM/K2-Horizon-MoVA-36B-A4B-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
RTX PRO 4500 (32 GB)CheapestCheapest7 GB free	~146 tok/s	$0.65	Deploy on RunPodfrom $0.34/hr (opens in new tab)	32 GB · 7 GB free · ~146 tok/s · $0.65 / 1M tokens
A40 (48 GB)23 GB free	~124 tok/s	$0.79	Deploy on RunPodfrom $0.35/hr (opens in new tab)	48 GB · 23 GB free · ~124 tok/s · $0.79 / 1M tokens
A40 (48 GB)23 GB free	~124 tok/s	$0.90	Deploy on Vast.aifrom $0.40/hr (opens in new tab)	48 GB · 23 GB free · ~124 tok/s · $0.90 / 1M tokens
RTX 6000 Ada (48 GB)23 GB free	~152 tok/s	$1.22	Deploy on Vast.aifrom $0.67/hr (opens in new tab)	48 GB · 23 GB free · ~152 tok/s · $1.22 / 1M tokens
RTX 5090 (32 GB)7 GB free	~211 tok/s	$0.91	Deploy on RunPodfrom $0.69/hr (opens in new tab)	32 GB · 7 GB free · ~211 tok/s · $0.91 / 1M tokens
RTX PRO 4500 Server (32 GB)7 GB free	~136 tok/s	$1.47	Deploy on RunPodfrom $0.72/hr (opens in new tab)	32 GB · 7 GB free · ~136 tok/s · $1.47 / 1M tokens
L40S (48 GB)23 GB free	~143 tok/s	$1.56	Deploy on Vast.aifrom $0.80/hr (opens in new tab)	48 GB · 23 GB free · ~143 tok/s · $1.56 / 1M tokens
RTX 5090 (32 GB)7 GB free	~211 tok/s	$1.10	Deploy on Vast.aifrom $0.83/hr (opens in new tab)	32 GB · 7 GB free · ~211 tok/s · $1.10 / 1M tokens

Compare GPUs →