Model: K2 Horizon MoVA 36B-A4B

MBZUAI IFM released K2 Horizon MoVA 36B-A4B, a 37.5B-parameter mixture-of-experts model with roughly 4B active parameters per token, under Apache 2.0 and with its training data. The model's 48 layers each keep an ordinary KV cache, so Q4_K_M needs about 24.2 GB of VRAM at 4K context — tight on a 24 GB card and comfortable on a 32 GB card up to 16K — while the full 512K window would require about 129 GB. IFM published four GGUF quant variants (Q4_K_M at 22.37 GB, Q5_K_M at 26.44 GB, Q6_K at 30.77 GB and Q8_0 at 39.83 GB), and llama.cpp runs the model as `k2-horizon`; no per-level quality loss has been published.

K2 Horizon MoVA 36B-A4B 36B-A4B MBZUAI K2 Horizon The mixture-of-experts member of MBZUAI IFM's K2 Horizon family: 37.5B parameters with about 4B active per token, Apache 2.0, released with its training data. Every one of its 48 layers keeps an ordinary KV cache, so context is not cheap: Q4 K M is about 24.2 GB at 4K, which is tight on a 24 GB card, and comfortable on a 32 GB card up to 16K. The 512K window would need about 129 GB. llama.cpp runs it as k2-horizon . Sizes come from IFM's own GGUF files bartowski's Q4 K M is about 0.3 GB larger ; no per-level quality loss has been published. 524K Max Context 4 Quant Variants GGUF Q8 0 Best Quality — Accuracy Retained Quantization Variants Per-quant VRAM, quality loss, and inference speed on RTX 4090 Measured = run on this site · Estimated = not run here: calculated, or a figure this site has not verified · Community = public reports | Format | Level | BPW | VRAM | PPL Loss | Speed | Source | Actions | |---|---|---|---|---|---|---|---| | GGUF | Q4 K M | 4.78 | 24.2 GB | — | — | Estimated | | | GGUF | Q5 K M | 5.65 | 28.5 GB | — | — | Estimated | | | GGUF | Q6 K | 6.57 | 33.0 GB | — | — | Estimated | | | GGUF | Q8 0 | 8.51 | 42.5 GB | — | — | Estimated | | Which file each size is for A file size is not a memory figure: the VRAM column adds the KV cache at 4K context and runtime buffers to the weights. A vision projector mmproj , where the repo ships one, is a separate download and is in neither number. - GGUF Q4 K M IFM/K2-Horizon-MoVA-36B-A4B-GGUF https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B-GGUF · K2-Horizon-MoVA-36B-A4B-Q4 K M.gguf · 22.37 GB 22,368,011,616 bytes · listing read 2026-10-21 - GGUF Q5 K M IFM/K2-Horizon-MoVA-36B-A4B-GGUF https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B-GGUF · K2-Horizon-MoVA-36B-A4B-Q5 K M.gguf · 26.44 GB 26,439,456,096 bytes · listing read 2026-10-21 - GGUF Q6 K IFM/K2-Horizon-MoVA-36B-A4B-GGUF https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B-GGUF · K2-Horizon-MoVA-36B-A4B-Q6 K.gguf · 30.77 GB 30,765,365,856 bytes · listing read 2026-10-21 - GGUF Q8 0 IFM/K2-Horizon-MoVA-36B-A4B-GGUF https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B-GGUF · K2-Horizon-MoVA-36B-A4B-Q8 0.gguf · 39.83 GB 39,831,174,496 bytes · listing read 2026-10-21 Similar models Compare with K2 Horizon https://quantized.uk/tools/compare/?a=k2-horizon-mova-36b-a4b&b=k2-horizon-32b 32BConsumer GPUPro GPU23.1 GBmin VRAM—accuracy https://quantized.uk/quant-hub/k2-horizon-32b/ K2 Horizon 32B MBZUAI K2 Horizon The dense 32B member of MBZUAI IFM's K2 Horizon family 34.8B counted, Apache 2.0, released with its training data . All 64 layers keep an ordinary KV cache with 8 KV heads: Q4 K M is about 23.1 GB at 4K, tight on a 24 GB card and comfortable on a 32 GB card up to 16K; each extra 32K of context adds roughly 8.8 GB. llama.cpp runs it as k2-horizon . Sizes come from IFM's own GGUF files bartowski's Q4 K M is about 1 GB larger ; no per-level quality loss has been published. 7BConsumer GPUMac / Apple Silicon6.5 GBmin VRAM—accuracy https://quantized.uk/quant-hub/k2-horizon-7b/ K2 Horizon 7B MBZUAI K2 Horizon The small dense member of MBZUAI IFM's K2 Horizon family, Apache 2.0, released with its training data. It is called 7B but holds 9.0B parameters: the 250K-token vocabulary's input and output tables add about 2B. Q4 K M is about 6.5 GB at 4K, comfortable on an 8 GB card at that length; a 16 GB card holds it comfortably to 32K 10.8 GB , and 128K needs about 25.6 GB because every layer keeps an ordinary KV cache. llama.cpp runs it as k2-horizon . Sizes come from IFM's own GGUF files bartowski's Q4 K M is about 0.2 GB larger ; no per-level quality loss has been published. 36BConsumer GPUPro GPU19.5 GBmin VRAM98.7%accuracy https://quantized.uk/quant-hub/seed-oss-36b/ Seed-OSS 36B Instruct ByteDance Seed Dense 36B with a native 512K context. The Q4 K M file is 21.8 GB, so with cache and buffers it needs about 23.9 GB at 4K: tight on a 24 GB card, where Q3 K M 19.5 GB is the comfortable choice, or a 2×16GB split. The AWQ build is 21.1 GB, not the 4-bit rate the name suggests. Context is the real cost: 512K of KV cache is roughly 128GB on its own, so budget context first and weights second. 27BConsumer GPUPro GPU15.3 GBmin VRAM—accuracy https://quantized.uk/quant-hub/qwen3-8-27b/ Qwen3.8 27B Alibaba Qwen3.8 Hybrid attention: only 16 of its 64 layers keep a KV cache, which is why a 27B model holds a 262K context in about 16GB of cache rather than 64GB. Sizing it as a conventional stack overstates the cache fourfold — the estimates here count the 16 and match published measurements at 8K, 32K and 262K. Quality loss per level has not been published. Running K2 Horizon MoVA 36B-A4B locally At Q4 K M and 4K of context, K2 Horizon MoVA 36B-A4B needs about 24.2 GB — 21.3 GB of weights, 0.75 GB of KV cache and a 2.2 GB activation buffer. The smallest cards in this index that clear that comfortably are 32 GB ones — 4 of them, from the RTX 5090 up — and 24 of the 78 cards here do. These are calculated figures, not measurements: the estimate stops counting a card as comfortable at 88% of its VRAM, which is roughly the room a desktop session needs. What longer context costs Going from 4K to 32K adds about 5.25 GB, taking the total to 30.0 GB. Weights do not move with context — only the KV cache does, and it grows linearly, so this is the number to watch when planning for long documents. The model's native window is 512K; holding all of it at Q4 K M would need about 129 GB. Which build to download This index only tracks K2 Horizon MoVA 36B-A4B in GGUF, across 4 levels Q4 K M, Q5 K M, Q6 K, Q8 0 . No per-level perplexity sweep has been published for this model, so the quality column is empty rather than estimated — within one model, more bits per weight is the only ordering the data supports. No throughput has been measured for this model on this site's hardware. GGUF runs on llama.cpp and Ollama across NVIDIA, AMD and Apple silicon; AWQ and GPTQ target vLLM on CUDA and ROCm; EXL2 is ExLlamaV2 — CUDA only, and now archived, so local use rather than serving. Common questions - How much VRAM does K2 Horizon MoVA 36B-A4B need? - About 24.2 GB at Q4 K M with 4K of context and batch 1, rising to roughly 30.0 GB at 32K. That figure is weights plus KV cache plus a 10% activation buffer, calculated from the model's architecture rather than measured on a card. - Will K2 Horizon MoVA 36B-A4B run on a 32GB GPU? - Yes — at Q4 K M and 4K context it needs about 24.2 GB, which leaves 7.8 GB spare on a RTX 5090, or on any of the other 32 GB cards here 4 in all . That is the smallest size in this index that clears it comfortably; 24 of 78 do. - Which quantization of K2 Horizon MoVA 36B-A4B should I use? - No published quality comparison exists for this model, so pick by footprint: Q4 K M is the level most people run, and a higher bits-per-weight level is more faithful. All 4 levels in the index are Q4 K M, Q5 K M, Q6 K, Q8 0. Note this is a mixture-of-experts model — all parameters must be resident even though only a fraction are active per token, so the memory cost follows the total, not the active count. Where this model fits Sized at Q4 K M with a 4K context window, smallest card first. Comfortable means the estimate uses at most 88% of the memory. - Runs comfortably on - RTX 509032GB · +7.8 https://quantized.uk/gpu/rtx-5090/ Arc Pro B70 32G32GB · +7.8 https://quantized.uk/gpu/arc-pro-b70-32g/ Radeon AI PRO R9700 32G32GB · +7.8 https://quantized.uk/gpu/radeon-ai-pro-r9700-32g/ Instinct MI100 32G32GB · +7.8 https://quantized.uk/gpu/instinct-mi100-32g/ A100 40G40GB · +15.8 https://quantized.uk/gpu/a100-40g/ - Tight but possible - RTX 309024GB https://quantized.uk/gpu/rtx-3090/ Radeon RX 7900 XTX24GB https://quantized.uk/gpu/radeon-rx-7900-xtx/ RTX 409024GB https://quantized.uk/gpu/rtx-4090/ - Ships in - GGUF https://quantized.uk/formats/gguf/ - Try it yourself - Size it in the calculator https://quantized.uk/tools/vram-calc/?model=k2-horizon-mova-36b-a4b&quant=Q4 K M Generate the run command https://quantized.uk/tools/cli-gen/?model=k2-horizon-mova-36b-a4b&quant=Q4 K M How to actually run this Deployment guides for this model and this class of hardware. This page's figures change when the model or the runtime does.Last updated 2026-10-23 RSS → /feed.xml https://quantized.uk/feed.xml