K2-Horizon-MoVA-36B-A4B (MLX, bf16)
Unquantized bf16 MLX conversion of IFM/K2-Horizon-MoVA-36B-A4B, converted from revision e5c131d. The weights are the original bf16 values, repacked for MLX (expert weights stacked for gather_mm). 70 GB on disk. For Apple silicon.
K2-Horizon-MoVA-36B-A4B is IFM's sparse K2-Horizon model: a 36B-parameter Mixture-of-Experts model with 4B active per token and 512K context. Its attention is Mixture-of-Values (MoVA), where the values come from 4 of 64 routed value experts.
Requirements
mlx-lm doesn't support the k2_horizon architecture yet. There's an open request: mlx-lm#1876. Until support lands, this repo ships the MLX model code (k2_horizon.py), which mlx-lm loads through the model_file entry in config.json. So the released mlx-lm works as-is:
pip install -U mlx-lm
Pass --trust-remote-code (or trust_remote_code=True). mlx-lm 0.31.3 loads the file without it, but newer versions require it. The file runs on your machine, so read it first if you like.
Once K2-Horizon support lands in a released mlx-lm, this repo will be updated to use it. If something breaks after that, open a discussion here asking for a re-upload.
How it was converted
mlx_lm.convert --dtype bfloat16 with the K2-Horizon model code described below. No quantization; the weights are bit-identical to the original checkpoint.
Memory
Peak 75 GB for a short prompt, so it needs a 128 GB Mac. That's under the default Metal working-set cap on 128 GB machines, so no sysctl iogpu.wired_limit_mb change is needed. The KV cache adds about 192 KB per token (about 24 GB at 128K tokens), so long contexts are tight; use the 8-bit version if you need them.
Conversion check
The MLX implementation was checked against IFM's PyTorch code (modeling_k2_horizon.py) in fp32 on the real weights, one layer at a time. All 48 layers match to a relative error of 1e-6 or better, and the next-token predictions agree at every position.
Smoke-tested after conversion with released mlx-lm 0.31.3 (mlx_lm.generate and mlx_lm.server):
17 * 23→391with correct reasoning- "capital of Australia" →
Canberra - A multi-turn server conversation that carries context correctly
On a Mac Studio M4 Max 128GB: 35.3 tok/s generation, peak 75.0 GB (short prompt).
Benchmarks (all K2-Horizon MLX variants)
WikiText-2 test perplexity (128 × 512 tokens, lower is better) and generation speed on an M4 Max 128GB, all measured the same way:
| bf16 | 8-bit | 4-bit | |
|---|---|---|---|
| Bits/weight | 16 | 8.5 | 5.6 |
| Disk | 70 GB | 37 GB | 25 GB |
| Peak memory | 75.0 GB | 39.9 GB | 26.5 GB |
| WikiText-2 perplexity | 11.368 | 11.372 (+0.04%) | 11.508 (+1.2%) |
| Generation | 35.3 tok/s | 51.2 tok/s | 59.6 tok/s |
Perplexity is a coarse signal. Test the versions on your own workload before picking one.
Usage
mlx_lm.generate --model mlx-community/K2-Horizon-MoVA-36B-A4B-bf16 --trust-remote-code --prompt "Explain mixture-of-experts in two sentences." --max-tokens 2048
from mlx_lm import load, generate
# Newer mlx-lm versions need trust_remote_code=True; on mlx-lm 0.31.3 use load("mlx-community/K2-Horizon-MoVA-36B-A4B-bf16").
model, tokenizer = load("mlx-community/K2-Horizon-MoVA-36B-A4B-bf16", trust_remote_code=True)
messages = [{"role": "user", "content": "Explain mixture-of-experts in two sentences."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tokenizer, prompt, max_tokens=2048))
mlx_lm.chat and mlx_lm.server take the same --trust-remote-code flag.
The model thinks before it answers, inside <ifm|think> … </ifm|think>. Leave room for that in max_tokens. The effort level is set with reasoning_effort in the chat template: "high" (default), "medium" or "low", e.g. apply_chat_template(messages, add_generation_prompt=True, reasoning_effort="low").
Notes:
- Server output: mlx-lm doesn't recognize the
<ifm|think>tags yet, somlx_lm.serverreturns the thinking text insidecontent, before</ifm|think>, rather than in a separatereasoningfield. K2-Horizon's tool-call format isn't parsed yet either. - Chat template change: the original template raises an error when an earlier assistant message has no thinking field, which is what OpenAI-style clients send. The template here renders empty thinking for those messages instead. Nothing else was changed.
License
Apache-2.0, inherited from the base model. Refer to the original model card for architecture, benchmarks and intended use. All credit for the model belongs to IFM.