K2-Horizon MoVA 36B-A4B

Does not run in stock llama.cpp, Ollama or LM Studio yet. Support is an open pull request (llama.cpp #29535, opened 2026-09-27 and flagged by the project's PR checker as too large to review as-is), so a stock build refuses the GGUF with "does not recognise this GGUF's model architecture". IFM's fork works, but early builds of it also failed to load on Windows over a tokenizer regex. On a Mac, the mlx-community build does run in stock mlx-lm via --trust-remote-code, though mlx_lm.server does not yet parse its thinking tags or tool calls.

In the one independent test, at Q5_K_M with a 256K context, a passkey planted at 0%, 25% or 50% depth was found three times out of three, but at 75% and 100% depth only once out of three, with noisy answers even when it succeeded. Fifteen runs on one setup, so a warning rather than a measurement; the HF thread that cites it summarises this as weak recall beyond 131K. The 512K context the card advertises should not be taken on trust.

The same test found it looping in long agentic coding tasks: it failed a Blender task and a Godot game build with repeated "thinking loop detected" interventions, kept running tests after being told to stop, and scored 57 of 100 on the reviewer's private HumanEval rewrite. The reviewer did not recommend it. IFM replied that it was not trained for those tasks and that improvements are coming. A separate thread reports the stock chat templates do not handle tool calls correctly in the Hermes agent.

Every one of its 48 layers keeps a full KV cache, which MoVA does not shrink: the routed value experts still write one value vector per KV head. That is 192KB per token, about 24GB at 128K, on top of 22GB of Q4_K_M weights. IFM acknowledges the attention layout "would take quite a bit memory for long generation" and says a fix is in the works.

IFM's benchmark table follows the Artificial Analysis Intelligence Index categories and fills every comparison column with Artificial Analysis data. This site does not republish Artificial Analysis numbers, so the AA-branded rows (AA-LCR, AA-Omniscience, tau3-Banking) are left out, and with them most of the evidence on tool calling, long context and factuality. Those cells are unrated, not bad. Absent from Aider polyglot and Vectara HHEM as of 2026-10-04.

IFM ships nothing below Q4_K_M; the 3-bit files are a community conversion (NANI-Nithin) that has not been measured. mlx-community found this model sensitive in its non-expert weights: plain 4-bit cost +4.5% perplexity, and keeping attention and embeddings at 8-bit brought that to +1.2%. Uniform 3-bit has nowhere to hide that, so the band is scored two points down.