UraionLabs/K2-Horizon-7B-oQ4e

Uraion Labs
_{Foundational systems research.}

K2-Horizon-7B — oQ4e

An MLX oQ4e mixed-precision quantization of IFM/K2-Horizon-7B for efficient local inference on Apple Silicon, quantized and published by Uraion Labs.

K2-Horizon-7B-oQ4e is an MLX mixed-precision quantization of IFM/K2-Horizon-7B, optimized for efficient local inference on Apple Silicon Macs.

This checkpoint was quantized and published by Uraion Labs using the oMLX oQe quantization workflow.

The oQ4e configuration targets a substantially smaller memory and storage footprint than higher-precision variants while retaining additional precision for model components identified as more quantization-sensitive by the oQe workflow.

The checkpoint uses the standard MLX model format and is intended primarily for oMLX and compatible MLX inference runtimes.

The underlying K2-Horizon architecture, tokenizer, chat template, reasoning behavior, tool-use format, and native 524,288-token context architecture originate from the upstream IFM model.

About This Quantization

• Quantization Level: oQ4e
• Quantization Workflow: oMLX oQe mixed-precision quantization
• Base Weight Precision: 4-bit
• Base Model: IFM/K2-Horizon-7B
• Model Class: 7B dense decoder-only
• Quantization Strategy: sensitivity-aware mixed precision
• Target Runtime: Apple Silicon macOS
• Storage Format: MLX safetensors
• Quantized By: Uraion Labs
• Native Context Architecture: 524,288 tokens
• Reasoning Support: inherited from upstream
• Tool Calling: inherited from upstream
• License: Apache-2.0

Exact layer overrides, effective bits per weight, group size, calibration statistics, output-head precision, and final repository size should be taken directly from the oMLX quantization report for this checkpoint.

Model Information

Property	Value
Original Model	IFM/K2-Horizon-7B
Quantization Variant	oQ4e
Quantized By	Uraion Labs
Quantization Tooling	oMLX
Quantization Workflow	oQe mixed precision
Architecture	Dense decoder-only language model
Model Class	7B
Base Weight Precision	4-bit
Mixed Precision	Yes
Context Length	524,288 tokens
Reasoning	K2-Horizon reasoning format
Tool Calling	K2-Horizon tool-call format
Storage Format	MLX safetensors
Target Platform	Apple Silicon macOS
Intended Runtime	omlx, MLX-compatible runtimes
License	Apache-2.0

Why oQ4e?

The oQ4e variant is intended to provide a stronger balance between:

• model quality,
• memory footprint,
• storage footprint,
• inference throughput,
• and practical local deployment.

Rather than uniformly quantizing every eligible model component to the same precision, the oQe workflow can retain additional precision for tensors that are more sensitive to quantization.

This makes oQ4e particularly suitable for users who want to run K2-Horizon-7B locally without the larger memory footprint of an 8-bit checkpoint.

The exact quality/performance trade-off should be measured against the final quantized artifact rather than inferred from nominal bit width alone.

Upstream K2-Horizon-7B Highlights

K2-Horizon-7B is developed by IFM as the medium dense member of the K2-Horizon family.

7B Dense Architecture

K2-Horizon-7B is a 7B-class dense decoder-only language model evaluated across:

• mathematical reasoning,
• software engineering,
• scientific reasoning,
• scientific coding,
• long-context reasoning,
• terminal use,
• tool use,
• search,
• and agentic workflows.

Native 512K Context

The model supports a native context architecture of:

524,288 tokens

IFM progressively extended context during midtraining through:

8K → 32K → 128K → 512K

Practical usable context during local inference depends on available unified memory, KV-cache configuration, runtime implementation, prompt size, and concurrent workload.

Agentic Capabilities

The upstream model was evaluated on benchmarks including:

• SWE-bench Verified
• Terminal-Bench 2.1
• tau3-Banking
• BrowseComp
• SciCode
• LCR
• HLE
• HMMT

Reasoning Control

K2-Horizon supports request-level reasoning effort through its chat template.

IFM recommends:

{
  "chat_template_kwargs": {
    "reasoning_effort": "high"
  }
}

Tool Calling

The upstream model supports structured tool-use formats including:

json
xml
xml_typed

The upstream default format is:

xml

Benchmark Results

The results below are reported by IFM for the original K2-Horizon-7B checkpoint. They are not independent benchmark results for this Uraion Labs oQ4e quantization.

Capability	Benchmark	K2-Horizon-7B	Reference Models
Math	HMMT Feb 2026	73.3	Gemma 4-12B: 63.1 · Qwen3.5-9B: 65.7 · Granite 4.2-8B: 66.5
Coding	SWE-bench Verified	70.6	Gemma 4-12B: 30.6 · Granite 4.2-8B: 47.7 · Qwen3.5-9B: 50.8
Scientific Reasoning	HLE	18.6	Granite 4.2-8B: 9.7 · Qwen3.5-9B: 14.9 · Gemma 4-12B: 15.7
Scientific Coding	SciCode	31.6	Qwen3.5-9B: 27.5 · Mistral Small 4: 28.0 · Granite 4.2-8B: 30.4
Long Context	LCR	68.0	Granite 4.2-8B: 43.3 · Gemma 4-12B: 61.7 · Qwen3.5-9B: 65.3
Agentic Coding	Terminal-Bench 2.1	39.1	Granite 4.2-8B: 18.4 · Gemma 4-12B: 27.3 · Qwen3.5-9B: 29.2
Tool Use	tau3-Banking	25.8	Qwen3.5-9B: 7.0 · Granite 4.2-8B: 7.6 · Muse Glimmer-30B: 24.0
Web Search	BrowseComp	59.0	DeepSeek V4 Flash-0423: 53.5 · GPT-5: 54.9 · LongCat Flash Thinking-2601: 56.6

For BrowseComp, IFM reports that K2-Horizon-7B uses the Discard-all@95k context-length protocol proposed in the DeepSeek-V3.2 technical report.

Comparison models may have been evaluated using different harnesses or inference configurations.

Quantization and Benchmark Disclaimer

Quantization changes numerical precision and may affect model behavior.

The upstream scores above therefore should not be interpreted as measured scores for:

K2-Horizon-7B-oQ4e

Possible differences may occur in:

• reasoning accuracy,
• code generation,
• tool-call formatting,
• long-context retrieval,
• instruction following,
• structured output,
• perplexity,
• and generation stability.

Independent benchmarking of the final oQ4e checkpoint is recommended when exact quality retention matters.

Quickstart

Using oMLX

oMLX is the primary intended inference runtime for this release on Apple Silicon.

Install oMLX:

pip install omlx

Run the model:

omlx run UraionLabs/K2-Horizon-7B-oQ4e \
  --prompt "Explain why long-context reasoning is difficult."

Hugging Face Download

Download the repository locally with:

hf download UraionLabs/K2-Horizon-7B-oQ4e

Using mlx-lm

If the final checkpoint is compatible with the installed version of mlx-lm, install it with:

pip install mlx-lm

Python Example

from mlx_lm import load, generate

model_id = "UraionLabs/K2-Horizon-7B-oQ4e"

model, tokenizer = load(model_id)

messages = [
    {
        "role": "user",
        "content": "Explain why long-context evaluation is difficult."
    }
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=32768,
    temp=1.0,
    top_p=0.95,
    verbose=True,
)

print(response)

Verify mlx-lm compatibility against the final checkpoint and runtime version. oMLX remains the intended runtime for this release.

OpenAI-Compatible Serving

When serving through an oMLX OpenAI-compatible endpoint, use:

UraionLabs/K2-Horizon-7B-oQ4e

For example, an OpenAI-compatible client can connect to the server exposed by your local oMLX deployment.

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8000/v1",
    api_key="omlx",
)

response = client.chat.completions.create(
    model="UraionLabs/K2-Horizon-7B-oQ4e",
    messages=[
        {
            "role": "user",
            "content": "Explain speculative decoding in simple terms."
        }
    ],
    temperature=1.0,
    top_p=0.95,
    max_tokens=32768,
)

print(response.choices[0].message.content)

The exact local URL, port, API key, and served model identifier depend on the user's oMLX configuration.

Recommended Generation Settings

IFM recommends the following configuration for the original K2-Horizon-7B model:

Parameter	Recommended Value	Notes
reasoning_effort	high	Used for upstream reported evaluations
temperature	1.0	Recommended sampling temperature
top_p	0.95	Nucleus sampling
Output Budget	>= 32,768 tokens	Avoid reasoning truncation

Recommended upstream sampling configuration:

{
  "temperature": 1.0,
  "top_p": 0.95,
  "max_tokens": 32768
}

Recommended reasoning configuration:

{
  "chat_template_kwargs": {
    "reasoning_effort": "high"
  }
}

Lower reasoning-effort settings may reduce latency and output length at the cost of capability.

Reasoning Format

K2-Horizon separates internal reasoning output and normal response content when used with runtimes that support its dedicated parser.

The upstream serving configuration uses:

k2_horizon

as the reasoning parser.

For example, upstream vLLM configurations use:

--reasoning-parser k2_horizon

Whether reasoning is exposed separately through oMLX depends on runtime support and API configuration.

Tool Calling

K2-Horizon supports multiple tool-call formats.

Supported upstream formats are:

json
xml
xml_typed

The default is:

xml

When using runtimes that implement the upstream K2 parser, IFM recommends:

--tool-call-parser k2_horizon

For oMLX deployments, preserve the original tokenizer and chat template wherever possible.

Do not substitute a generic chat template unless the replacement has been specifically validated for:

• reasoning,
• tool calling,
• multi-turn conversations,
• structured output,
• and agentic workloads.

Upstream vLLM Example

The original BF16 checkpoint can be served using IFM's recommended vLLM configuration:

vllm serve IFM/K2-Horizon-7B \
  --trust-remote-code \
  --dtype bfloat16 \
  --tensor-parallel-size 1 \
  --reasoning-parser k2_horizon \
  --enable-auto-tool-choice \
  --tool-call-parser k2_horizon

This command applies to the original upstream checkpoint, not this MLX quantization.

It is included to document the model's intended reasoning and tool-calling configuration.

Upstream SGLang Example

IFM also provides a validated SGLang configuration for the original model:

sglang serve \
  --model-path IFM/K2-Horizon-7B \
  --revision 69ada542b68fe13d767479db2ab9421baff88681 \
  --tp 1 \
  --dtype bfloat16 \
  --attention-backend fa3 \
  --reasoning-parser k2_horizon \
  --host 0.0.0.0 \
  --port 30000

Again, this is provided for upstream reference and does not apply directly to the MLX-formatted oQ4e release.

Training Overview

Uraion Labs did not train K2-Horizon-7B.

The following training information describes the upstream IFM model.

K2-Horizon-7B was produced through a multi-stage curriculum consisting of:

1. large-scale pretraining,
2. progressive long-context extension,
3. specialist reinforcement learning,
4. expert merging,
5. supervised fine-tuning.

Upstream Training Stages

Training Stage	Steps	Additional Tokens	Sequence Length	Purpose
Pretraining	1,100,000	22.9T	8K	General pretraining
Midtraining — Stage 1	55,000	1.1T	32K	Context extension
Midtraining — Stage 2	25,000	498B	128K	Context extension
Midtraining — Stage 3	5,500	110B	512K	Context extension
Midtraining — Stage 4	10,000	199B	512K	Agentic + reasoning-oriented continuation
RL — Math Expert	2,399	29.1B	64K	Mathematical reasoning RL
RL — Code Expert Stage 1	601	6.2B	64K	Code RL
RL — Code Expert Stage 2	1,499	12.3B	64K	Broader code RL
RL — Search Expert	59	8.4B	—	Search-agent RL
RL — Tool-Use Expert	39	1.4B	—	Tool-use RL
RL — Merge	—	—	—	Specialist-model merge
SFT — Phase 1	10,000	199B	512K	Domain coverage
SFT — Phase 2	2,500	50B	512K	High-quality SFT + LR decay

Progressive Context Extension

The original model's context length was extended progressively during midtraining:

Pretraining
8K
 ↓
Midtraining Stage 1
32K
 ↓
Midtraining Stage 2
128K
 ↓
Midtraining Stage 3
512K
 ↓
Midtraining Stage 4
512K

This differs from simply modifying a RoPE or runtime context parameter after training: the model received additional training at progressively longer sequence lengths.

Reinforcement Learning Experts

During upstream RL, IFM branched training into specialized models.

These included:

• Math expert
• Code expert
• Search expert
• Tool-use expert

The final RL model was produced using a merge that incorporated:

Midtraining Stage 4
        +
Math Expert
        +
Code Expert Stage 2
        +
Search Expert
        +
Tool-Use Expert
        ↓
   RL Merged Model
        ↓
    SFT Phase 1
        ↓
    SFT Phase 2

The released K2-Horizon-7B checkpoint follows this broader training lineage.

Training Data

Training-data metadata is inherited from the upstream IFM release.

Referenced upstream datasets include:

• IFM/K2-Horizon-Pretrain-Data
• IFM/K2-Horizon-Midtrain-Data

Uraion Labs did not create or modify these datasets as part of this quantization.

This repository represents a quantized derivative of the released model weights rather than a retrained model.

Quantization vs. Training

The roles of IFM and Uraion Labs are distinct.

IFM

IFM:

• designed K2-Horizon,
• trained K2-Horizon-7B,
• performed pretraining,
• performed long-context midtraining,
• performed reinforcement learning,
• trained specialist models,
• merged the RL experts,
• performed supervised fine-tuning,
• produced the original model weights,
• and reported the upstream benchmark results.

Uraion Labs

Uraion Labs:

• started from the released IFM/K2-Horizon-7B checkpoint,
• converted the checkpoint for MLX deployment,
• quantized the model using oMLX,
• used the oQe mixed-precision workflow,
• produced this oQ4e variant,
• and published the Apple Silicon-oriented derivative.

Uraion Labs did not retrain the underlying model as part of this release.

No capability improvement over the upstream model should be inferred from the quantization itself.

Quantization Metadata

Known metadata for this release:

Property	Value
Quantization	oQ4e
Tooling	oMLX
Workflow	oQe
Base Precision	4-bit
Allocation	Mixed precision
Platform	Apple Silicon
Format	MLX safetensors
Original Model	IFM/K2-Horizon-7B
Quantized By	Uraion Labs

The following values should be inserted directly from the final oMLX quantization report:

Property	Value
Exact Mixed-Precision Profile	TBD from oMLX report
5-bit Layer Overrides	TBD from oMLX report
6-bit Layer Overrides	TBD from oMLX report
8-bit Layer Overrides	TBD from oMLX report
lm_head Precision	TBD from oMLX report
Group Size	TBD from oMLX report
Quantization Mode	TBD from oMLX report
Non-Quantized Tensor Precision	TBD from oMLX report
Calibration Dataset	TBD from oMLX report
Calibration Sample Count	TBD from oMLX report
Model Weight Size	TBD from final artifact
Effective Bits per Weight	TBD from oMLX report

These values are deliberately not inferred from other oQ4e releases because oQe's mixed-precision allocation can differ between model architectures and calibration runs.

Practical Context Length

Although the architecture supports:

524,288 tokens

the maximum practical context on a particular Mac depends heavily on:

• total unified memory,
• memory available to the GPU,
• model weight footprint,
• KV-cache precision,
• KV-cache implementation,
• runtime overhead,
• number of concurrent requests,
• prompt length,
• output reservation,
• and macOS memory pressure.

Quantizing model weights reduces the static weight footprint but does not eliminate the memory cost of very large KV caches.

Users should therefore distinguish between:

architectural context support

and

context that can practically be served within available memory.

Intended Use

This quantization is intended for:

• local LLM inference,
• Apple Silicon deployments,
• coding assistants,
• agent runtimes,
• tool-calling applications,
• long-context experimentation,
• local search agents,
• terminal agents,
• reasoning workloads,
• development environments,
• research,
• and edge deployment experiments.

Limitations

Quantization Loss

The oQ4e checkpoint uses substantially reduced numerical precision compared with the original BF16 model.

This can introduce quality degradation.

The amount of degradation may vary across capabilities.

Potentially sensitive workloads include:

• complex mathematical reasoning,
• long chains of reasoning,
• long-context retrieval,
• exact structured generation,
• software engineering,
• tool-call syntax,
• rare-token generation,
• and multilingual output.

Benchmark Transfer

The benchmark scores shown in this model card belong to the upstream IFM checkpoint.

They should not be treated as measured oQ4e results.

Runtime Differences

Generation behavior can differ depending on:

• runtime,
• sampler implementation,
• chat-template implementation,
• reasoning-parser support,
• context handling,
• quantized kernels,
• and generation parameters.

Hallucination

This model can produce:

• incorrect facts,
• unsupported claims,
• invalid code,
• incorrect tool calls,
• fabricated citations,
• and misleading reasoning.

Outputs should be independently verified when correctness matters.

Disclaimer

This model generates text based on statistical language-modeling patterns.

Responses involving specialized or consequential domains should not be treated as a substitute for qualified professional judgment.

Users are responsible for:

• validating outputs,
• implementing appropriate guardrails,
• testing the model for their intended application,
• and complying with applicable laws, licenses, regulations, and policies.

The model is provided AS IS, without warranty of any kind, express or implied.

Original Model and Attribution

This repository contains an MLX quantized derivative of:

IFM/K2-Horizon-7B

K2-Horizon-7B was developed and trained by IFM.

The following originate from the upstream K2-Horizon release:

• architecture,
• tokenizer,
• model training,
• pretraining data,
• midtraining,
• long-context training,
• reinforcement learning,
• supervised fine-tuning,
• reasoning behavior,
• tool-use behavior,
• benchmark results,
• and original model weights.

Uraion Labs performed the MLX/oMLX quantization and publication of this oQ4e derivative.

Upstream Resources

• Original Model: IFM/K2-Horizon-7B
• K2 Horizon Blog: Introducing K2 Horizon
• Training Logs: Weights & Biases
• Training Code: ifm-ai/xllm
• SGLang Cookbook: K2-Horizon

License

The underlying K2-Horizon-7B model is released under the Apache License 2.0.

This quantized derivative retains the applicable upstream licensing terms.

Users should review the original model repository and license before redistribution or deployment.

Citation

Please cite the original K2-Horizon work when referencing the underlying model:

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

_{Quantized and published by Uraion Labs using oMLX · Foundational systems research.}