---
base_model: IFM/K2-Horizon-MoVA-36B-A4B
tags:
  - mlx
  - apple-silicon
  - text-generation
  - oQ
license: apache-2.0
---

# K2-Horizon-MoVA-36B-A4B oQ4e

oQ4e (imatrix-enhanced mixed-precision 4-bit) conversion of [IFM/K2-Horizon-MoVA-36B-A4B](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B), a sparse Mixture-of-Experts model with Mixture-of-Values attention (36B total / 4B active parameters, 512K context).

**Upstream model:** [IFM/K2-Horizon-MoVA-36B-A4B](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B) by the IFM Team, released under Apache-2.0.

**Conversion:** Quantized to MLX format using [Hermes Agent](https://hermes-agent.nousresearch.com) with `mlx-lm` and `oMLX`.

## Quickstart

```bash
pip install -U mlx-lm

python3 -m mlx_lm.generate \
  --model hermitdave/K2-Horizon-MoVA-36B-A4B-oQ4e \
  --prompt "Explain why long-context evaluation is difficult." \
  --max-tokens 512 --temp 1.0 --top-p 0.95
```

## Reasoning

K2-Horizon is a reasoning model. Always use `reasoning_effort="high"` for best results:

```python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
    model="hermitdave/K2-Horizon-MoVA-36B-A4B-oQ4e",
    messages=[{"role": "user", "content": "Explain quantum entanglement."}],
    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)
```

## Benchmark Results

### Leaderboard Benchmarks

| Benchmark | K2-Horizon-MoVA-36B-A4B |
|-----------|------------------------|
| tau3-Banking (Agentic tool use) | **26.8** |
| Terminal-Bench 2.1 (Agentic terminal use) | **58.6** |
| GPQA Diamond (Graduate-level science QA) | 80.8 |
| AA-LCR (Long-context reasoning) | 66.3 |

Scores in %. See [model card](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B) for full results.

### Head-to-Head Comparison vs Qwen 3.6 35B A3B

Independent benchmark comparison — raw scores collected via [BenchLocal](https://github.com/hermitdave/benchlocal), report compiled by Hermes Agent. See [full report](https://hermitdave.github.io/benchmarks/local-moa-report.html).

| Benchmark | K2 Score | Qwen Score | Delta | Winner |
|-----------|----------|------------|-------|--------|
| Bugfix / Coding | 76 | 73 | +3 | K2 |
| Information Extraction | 90 | 84 | +6 | K2 |
| Formatting & Structured Output | 93 | 92 | +1 | K2 |
| Prompt Authority / Safety | 80 | 60 | +20 | K2 |
| Reasoning & Maths | 78 | 80 | -2 | Qwen |
| Structured Output | 95 | 87 | +8 | K2 |
| Tool Call Performance | 93 | 63 | +30 | K2 |
| Hermes Agent Capabilities | 54 | 48 | +6 | K2 |
| **Average** | **82.4** | **73.4** | **+9.0** | **K2** |

**K2 wins 7/8 benchmarks.** Dominates extraction, formatting, tool calling, and safety. Qwen wins only Reasoning & Maths by 2 points.

### Throughput Comparison (oMLX)

| Metric | K2 Horizon 36B A4B | Qwen 3.6 35B A3B | Delta |
|--------|-------------------|-----------------|-------|
| TG TPS (single stream) | 48.1 | 88.4 | +84% Qwen |
| TTFT (ms) | 6,300 | 2,371 | -62% Qwen |
| TPOT (ms) | 20.9 | 11.4 | -46% Qwen |
| Peak Memory (GB) | 22.0 | 21.9 | Tie |

**Qwen is ~1.8x faster** due to 3B active parameters (vs 4B) + MTP (multi-token prediction). K2 trades throughput for quality.

### Verdict by Dimension

| Dimension | Winner | Notes |
|-----------|--------|-------|
| **Quality** | K2 Horizon | Wins 7/8 benchmarks, +9.0 avg |
| **Throughput** | Qwen 3.6 | ~1.8x faster single-stream and batch |
| **Safety** | K2 Horizon | 0 harmful violations vs 1 |
| **Agentic Use** | K2 Horizon | Tool call: 93 vs 63, Error recovery: 100 vs 33 |
| **Multimodal** | Qwen 3.6 | Native text + image + video |
| **Long Context** | Draw | Qwen 1M (ext.) vs K2 512K native |

**Bottom line:** For local agent pipelines where tool reliability and structured output matter: **K2 Horizon**. For high-volume batch inference where throughput dominates: **Qwen 3.6**.

## oMLX Patch

K2-Horizon requires oMLX v0.6.4+ with the [K2-Horizon support patch (PR #3441)](https://github.com/jundot/omlx/pull/3441). This patch adds:

- `k2_horizon` model type support
- Reasoning content handling (`<ifm|think>` tags)
- Tool call parsing (plain text and XML formats)
- Multi-turn conversation support

Without this patch, oMLX will refuse to load K2-Horizon models with `ValueError: Model type k2_horizon not supported`.

## Chat Template

K2-Horizon uses IFM's custom chat template with reasoning and tool calling support. Key tags:

| Tag | Purpose |
|-----|---------|
| `<ifm\|think>`, `<ifm\|think_fast>`, `<ifm\|think_faster>` | Thinking blocks |
| `<\|ifm\|im_start|>`, `<\|ifm\|im_end\|>` | Message delimiters |
| `<ifm\|tool_call>`, `<ifm\|arg_key>`, `<ifm\|arg_value>` | Tool call structure |

All tags are automatically stripped by oMLX before responses reach users.

## Citation

```bibtex
@misc{k2horizon2026,
  title  = {Introducing K2 Horizon: Frontier Performance, Radically Open},
  author = {{IFM Team}},
  year   = {2026},
  url    = {https://ifm.ai/blog/k2/},
}
```

## License

Apache-2.0 (same as upstream).