Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.
K2 Horizon Fits on Paper. Not a GX10 Sit.
IFM K2 Horizon landed Sep 3, 2026 (ifm.ai/blog/k2/). Six sizes: 375B-A23B (~23B active), MoVA 36B-A4B, dense 32B, 7B, 3.7B, 0.9B. Cards and GGUFs are license:apache-2.0. Org huggingface.co/IFM.
Key takeaways
- IFM K2 Horizon landed Sep 3, 2026 (ifm.ai/blog/k2/). Six sizes: 375B-A23B (~23B active), MoVA 36B-A4B, dense 32B, 7B, 3.7B, 0.9B. Cards and GGUFs are license:apache-2.0. Org huggingface.co/IFM.
- Sit-size instruments are the GGUF LFS files and FP8 usedStorage — not main-repo usedStorage (those include checkpoint trees). MoVA BF16 GGUF 74,924,627,296; 32B 69,571,404,896; 7B 18,010,413,440; 3.7B file K2-Horizon-4B-BF16.gguf 10,128,343,424; 0.9B file K2-Horizon-1B-BF16.gguf 2,159,424,896. FP8 usedStorage: MoVA 48,387,791,853; 32B 37,372,747,849; 375B-A23B 390,586,256,132.
- Under-~110 GB siblings fit on paper. 375B FP8 does not. Card: 32B is Stage1 (final checkpoint to be released); MoVA says final checkpoint released. MoVA architecture claim is theirs. Stock llama.cpp has no k2_horizon as of b10786; GGUF card points at an IFM fork / PR in progress. ollama.com/library/k2-horizon is HTTP 404. I did not load this. Not a bench. Not a GX10 sit.
Local LLMs on NVIDIA Spark / ASUS GX10
IFM posted K2 Horizon this morning. The instrument is the IFM blog plus the Hugging Face cards under IFM. Fleet: 375B-A23B (~23B active), MoVA 36B-A4B, dense 32B, 7B, 3.7B, 0.9B. License on the cards I pulled: apache-2.0. Blog says the same. I am not rewriting another post for this.
Do not use main-repo usedStorage as the sit size. Those trees carry intermediate checkpoints. MoVA main is about 2.47 TB of usedStorage. 32B main is about 1.12 TB. 375B main is about 29.4 TB. That is the development tree, not the inference sibling.
The sit-size receipts are the official GGUF LFS bytes and the FP8 usedStorage values:
IFM/K2-Horizon-MoVA-36B-A4B-GGUF→K2-Horizon-36B-BF16.gguf74,924,627,296IFM/K2-Horizon-32B-GGUF→K2-Horizon-32B-BF16.gguf69,571,404,896IFM/K2-Horizon-7B-GGUF→K2-Horizon-7B-BF16.gguf18,010,413,440IFM/K2-Horizon-3.7B-GGUF→ filenameK2-Horizon-4B-BF16.gguf10,128,343,424IFM/K2-Horizon-0.9B-GGUF→ filenameK2-Horizon-1B-BF16.gguf2,159,424,896IFM/K2-Horizon-MoVA-36B-A4B-FP8usedStorage 48,387,791,853IFM/K2-Horizon-32B-FP8usedStorage 37,372,747,849IFM/K2-Horizon-375B-A23B-FP8usedStorage 390,586,256,132
Under-~110 GB GGUF/FP8 siblings fit on paper. 375B FP8 does not. That is a file-size claim, not a sit, not leftover unified memory, and not tokens per second.
Card claims I am not laundering into my own: MoVA is their Mixture-of-Value Attention story (~4B active). Native context on the cards is 512K (524,288). The 32B card is labeled Stage1 — "final checkpoint to be released." The MoVA card says they have released the final checkpoint. Those sentences are theirs.
Serve notes on the cards only: vLLM/SGLang recipes show --tensor-parallel-size 2 / --tp 2 for MoVA and 32B, and TP=8 for 375B-A23B. I am not standing those up. I am not quoting their H200 cookbook numbers.
Stock llama.cpp does not carry k2_horizon as of tag b10786. The official MoVA GGUF README says the PR is in progress and points at MBZUAI-IFM/llama.cpp branch model/K2Horizon. ollama.com/library/k2-horizon is HTTP 404 as of this check, even though the blog names day-zero Ollama support. Community GGUFs showed up the same morning; that is a separate line, not this receipt.
I run Grok when it earns it, and a GX10 when I want the weights in the room. The small and mid siblings sit on the local side of that split as downloads. The 375B FP8 does not. I have not loaded K2 Horizon on the GX10. If I sit a sibling and actually measure it, that note comes next. Not this card.
- Introducing K2 Horizon (IFM blog)
- IFM on Hugging Face
- IFM/K2-Horizon-MoVA-36B-A4B
- IFM/K2-Horizon-MoVA-36B-A4B-GGUF
- IFM/K2-Horizon-MoVA-36B-A4B-FP8
- IFM/K2-Horizon-32B
- IFM/K2-Horizon-32B-GGUF
- IFM/K2-Horizon-32B-FP8
- IFM/K2-Horizon-375B-A23B
- IFM/K2-Horizon-375B-A23B-FP8
- IFM/K2-Horizon-7B-GGUF
- IFM/K2-Horizon-3.7B-GGUF
- IFM/K2-Horizon-0.9B-GGUF
Frequently asked questions
Yes. IFM's blog is dated Sep 3, 2026. Hugging Face org IFM hosts the fleet, official GGUFs, and FP8 siblings under Apache 2.0.
On paper, the official under-~110 GB GGUF BF16 and FP8 siblings do as a file-size claim. Official 375B-A23B-FP8 usedStorage is 390,586,256,132 bytes and does not. Fit is not a sit. I have not loaded any of these on the GX10, and I am not inventing leftover VRAM or tokens per second.
Not from what I checked. ggml-org/llama.cpp src/llama-arch.cpp at tag b10786 has no k2_horizon. The official MoVA GGUF card says architecture support is a PR in progress and points at MBZUAI-IFM/llama.cpp branch model/K2Horizon. ollama.com/library/k2-horizon returns HTTP 404.
Affiliate Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.