mlx-community/K2-Horizon-3.7B-8bit

K2-Horizon-3.7B (MLX, 8-bit)

8-bit MLX quantization of IFM/K2-Horizon-3.7B, converted from revision 943ce4e. 8.5 bits/weight effective, 5.4 GB on disk. For Apple silicon.

K2-Horizon-3.7B is IFM's small dense K2-Horizon model: a 3.7B decoder-only model with a 512K (524,288-token) context window.

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; its comments describe how it differs from IFM's PyTorch code.

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 quantized

mlx_lm.convert -q --q-bits 8 --q-group-size 64 (integer affine quantization, bf16 scales and biases). All linear layers and the embeddings are 8-bit. WikiText-2 perplexity is within -0.05% of bf16, so it is effectively lossless.

Memory

Peak 5.5 GB for a short prompt; fits a 16 GB Mac. The KV cache adds about 144 KB per token in bf16 (18.0 GB at 128K tokens), so long contexts need more memory; --max-kv-size and KV-cache quantization (--kv-bits 8, where available) reduce it.

Conversion check

The MLX implementation was checked against IFM's PyTorch code (modeling_k2_horizon.py) in fp32 on the real weights of this model:

• Layer by layer, all 36 layers match to a relative error of 4e-6 or better, and the next-token predictions agree at every position.
• Token by token: the full model in fp32 greedily generated with the KV cache, and the PyTorch model picked the same token at every step (256/256 tokens across English, code, math and Chinese prompts). Cached and uncached outputs also agree at every step.

Smoke-tested after conversion with released mlx-lm 0.31.3: 17 * 23 → 391 and "capital of Australia" → Canberra, both ending normally. On a Mac Studio M4 Max 128GB: 84.2 tok/s generation, peak 5.5 GB (short prompt).

Benchmarks (all K2-Horizon-3.7B 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	6.5
Disk	10.1 GB	5.4 GB	4.1 GB
Peak memory	10.2 GB	5.5 GB	4.3 GB
WikiText-2 perplexity	17.552	17.544 (-0.05%)	18.324 (+4.4%)
Generation	50.3 tok/s	84.2 tok/s	109.6 tok/s

Perplexity is a coarse signal. Test the versions on your own workload before picking one. Other K2-Horizon sizes: the K2-Horizon collection.

Usage

mlx_lm.generate --model mlx-community/K2-Horizon-3.7B-8bit --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-3.7B-8bit").
model, tokenizer = load("mlx-community/K2-Horizon-3.7B-8bit", 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, so mlx_lm.server returns the thinking text inside content, before </ifm|think>, rather than in a separate reasoning field. 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.