K2-Horizon-7B-Uno Merged

Merged version of IFM/K2-Horizon-7B-Uno, a diffusion-augmented LLM based on K2-Horizon-7B. This repository contains the LoRA adapter merged directly into the base model weights using PyTorch + PEFT, so it runs as a standard autoregressive model without needing the custom Ψ-Spec sampler.

Upstream model: IFM/K2-Horizon-7B-Uno by Institute of Foundation Models, released under Apache 2.0.

Base model: IFM/K2-Horizon-7B

Conversion: Merged using PyTorch + PEFT, validated with text generation, then quantized to MLX format using Hermes Agent.

Architecture

K2-Horizon-7B-Uno combines:

AR pathway: Standard K2-Horizon-7B weights (next-token prediction)
Diffusion pathway: LoRA-based diffusion adapters (merged into AR weights)

Spec	Value
Total params	~7.6B (with merged LoRA)
Architecture	Dense + LoRA-augmented
Layers	32
Context length	512K native

Quickstart

pip install -U mlx-lm

python3 -m mlx_lm.generate \
  --model hermitdave/K2-Horizon-7B-Uno-merged \
  --prompt "Explain why long-context evaluation is difficult." \
  --max-tokens 512 --temp 1.0 --top-p 0.95

Reasoning

K2-Horizon-7B is a reasoning model. Always use reasoning_effort="high":

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
    model="hermitdave/K2-Horizon-7B-Uno-merged",
    messages=[{"role": "user", "content": "Explain step by step."}],
    extra_body={"chat_template_kwargs": {"reasoning_effort": "high"}},
)
print("Reasoning:", getattr(response.choices[0].message, "reasoning_content", None))
print("Answer:", response.choices[0].message.content)

oMLX Patch

K2-Horizon requires oMLX v0.6.4+ with the K2-Horizon support patch:

Registers k2_horizon model type
Recognizes IFM thinking tags (<ifm|think>, <ifm|think_fast>, <ifm|think_faster>)
Parses IFM tool call format (<ifm|tool_call>, <ifm|arg_key>, <ifm|arg_value>)
Strips tags in streaming and non-streaming paths

Citation

@misc{k2_horizon_7b_uno,
  title        = {K2-Horizon-7B-Uno},
  author       = {Institute of Foundation Models},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/IFM/K2-Horizon-7B-Uno}},
}

License

Apache 2.0 (same as upstream).

hermitdave/K2-Horizon-7B-Uno-merged