mlx-community/K2-Horizon-0.9B-8bit

K2-Horizon-0.9B (MLX, 8-bit)

8-bit MLX quantization of IFM/K2-Horizon-0.9B, converted from revision fa7f5de. 8.5 bits/weight effective, 1.2 GB on disk. For Apple silicon.

K2-Horizon-0.9B is IFM's compact dense K2-Horizon model: a 0.9B-class decoder-only model distilled from domain teachers, with a 128K (131,072-token) context window through YaRN RoPE scaling.

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 +0.3% of bf16, so it is effectively lossless.

Memory

Peak 1.3 GB for a short prompt; fits a 8 GB Mac. The KV cache adds about 56 KB per token in bf16 (7.0 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 28 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 (320/320 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: 306.9 tok/s generation, peak 1.3 GB (short prompt).

Benchmarks (all K2-Horizon-0.9B 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	6.4
Disk	2.2 GB	1.2 GB	0.9 GB
Peak memory	2.2 GB	1.3 GB	1.0 GB
WikiText-2 perplexity	18.654	18.713 (+0.3%)	19.703 (+5.6%)
Generation	195.7 tok/s	306.9 tok/s	379.8 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-0.9B-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-0.9B-8bit").
model, tokenizer = load("mlx-community/K2-Horizon-0.9B-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.
• End-of-turn token: the original generation_config.json lists only <|endoftext|> (id 1) as EOS, but the chat template ends each turn with <|ifm|im_end|> (id 64019). This repo adds 64019 to eos_token_id, like the other K2-Horizon models, so generation stops at the end of the answer and <|ifm|im_end|> does not appear in the output.
• dtype: the original config says float32, but the checkpoint weights are bf16. The config here says bfloat16, so no tool upcasts the weights.
• YaRN: RoPE uses YaRN (factor 16, original 8K context). The MLX code checks that the YaRN settings match the reference and raises an error for settings it does not support.
• Recommended sampling (from IFM's model card): temperature=0.6, top_p=0.95, and a large output budget for thinking.

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.