IFM/K2-Horizon-32B-NVFP4

K2-Horizon-32B-Stage1-NVFP4

Pre-training Code - Post-training Code - Pretraining Data - Midtraining Data

This repository contains an NVFP4-quantized version of IFM/K2-Horizon-32B.

All linear layers except lm_head are quantized to NVFP4:

• Weights: NVFP4
• Activations: NVFP4

The NVFP4 model shows slightly lower performance than the original BF16 model on our evaluations, while reducing memory footprint and enabling faster inference on NVFP4-capable hardware.

Serving note: Requires NVIDIA Blackwell-generation GPUs (B-series) or newer with native NVFP4 support.

K2-Horizon-32B-Stage1 is the large dense member of the K2-Horizon family: a 32B decoder-only model with a 512K context window. Note: final checkpoint to be released.

K2-Horizon-32B-Stage1 Highlights

• Strong dense baseline. A 32B dense model evaluated on the same agentic, coding, and reasoning benchmarks as the rest of the family (see Benchmark Results). Results are for stage 1 of the final model training; results for stage 2 will be out soon.
• 512K context. Native 524,288-token context from the midtraining stages onward.
• Intermediate checkpoints. Intermediate checkpoints will be released so capability changes can be studied across training rather than at a single checkpoint.
• Fully open. Training data/recipe and the training code will be made public.

Benchmark Results

		Open-weight dense models
	K2-Horizon-32B-Stage1	Qwen3.8-27B	Muse Glimmer-30B	IBM Granite 4.2 30B
# Params	32B	27B	30B	30B
# Activated params	32B	27B	30B	30B
Architecture	Dense	Dense	Dense	Dense
Agents
tau3-Banking Agentic tool use	22.5	48.0	23.5	14.4
Coding
Terminal-Bench 2.1 Agentic terminal use	36.6	79.8	51.7	26.6
SciCode Scientific coding	30.2	44.7	43.6	36.6
Scientific Reasoning
Humanity's Last Exam (without tools) Expert-level reasoning	22.8	33.9	22.0	11.2
GPQA Diamond Graduate-level science QA	82.3	90.5	83.5	64.4
CritPt Frontier physics reasoning	1.4	5.4	2.6	0.3
General
AA-LCR Long-context reasoning	65.3	77.3	80.0	46.7
AA-Omniscience Accuracy Factual accuracy	16.8	15.6	27.0	10.1
AA-Omniscience Non-Hallucination Non-hallucination rate	58.3	69.7	18.1	74.4

Scores in %. Bold marks the best score in each row. Sections follow the Artificial Analysis Intelligence Index categories. Baseline scores are from Artificial Analysis; Muse Glimmer-30B at high reasoning effort, other open models in their reasoning mode.

NVFP4 vs. BF16

K2-Horizon-32B-Stage1	IFEval (Prompt)	GSM8K	MBPP	MMLU-Pro	GPQA-Diamond	BBH (3-shot)	AIME 26 (avg @ 32)	Average
BF16	86.69	96.21	94.40	81.52	81.76	93.20	92.60	89.5
NVFP4	85.40	96.59	93.20	80.47	78.82	93.30	91.15	88.4

The evaluation context length is set to 65,536 tokens. Unless otherwise specified, all tasks are evaluated in a 0-shot setting. The NVFP4 models have currently been evaluated only on non-agent tasks. Results for agent tasks will be released later.

Quickstart

Serving

vLLM, recipe at recipes.vllm.ai/IFM:

vllm serve IFM/K2-Horizon-32B \
  --revision main \
  --model-impl vllm \
  --tensor-parallel-size 2 \
  --trust-remote-code \
  --dtype bfloat16 \
  --max-model-len 131072 \
  --reasoning-parser k2_horizon \
  --enable-auto-tool-choice \
  --tool-call-parser k2_horizon

SGLang recipe validated on 2× H200 in the SGLang K2 Horizon cookbook:

python3 -m sglang.launch_server \
  --model-path IFM/K2-Horizon-32B \
  --revision main \
  --tp 2 \
  --dtype bfloat16 \
  --attention-backend fa3 \
  --reasoning-parser k2_horizon \
  --tool-call-parser k2_horizon \
  --host 0.0.0.0 --port 30000

API Usage

Recommended settings: reasoning_effort="high", temperature=1.0, top_p=0.95. Reasoning depth is selected per request through chat_template_kwargs. Thinking is returned in reasoning_content and the answer in content.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
response = client.chat.completions.create(
    model="IFM/K2-Horizon-32B",
    messages=[{"role": "user", "content": "Explain the result step by step."}],
    temperature=1.0,
    top_p=0.95,
    max_tokens=32768,
    extra_body={"chat_template_kwargs": {"reasoning_effort": "high", "tool_call_format": "xml"}},
)
message = response.choices[0].message
print("Reasoning:", getattr(message, "reasoning_content", None))
print("Answer:", message.content)

Our model supports multiple tool calls formats, which can be changed with chat_template_kwargs. The supported values are json, xml, and xml_typed . The default is xml. Keep --tool-call-parser k2_horizon enabled to parse the selected format.

Transformers

Validated with Transformers 5.15.0, PyTorch 2.13.0, Safetensors 0.8.0.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "IFM/K2-Horizon-32B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, device_map="auto", dtype="bfloat16", low_cpu_mem_usage=True, trust_remote_code=True
)

inputs = tokenizer("Explain why long-context evaluation is difficult.", return_tensors="pt").to(model.device)
inputs.pop("token_type_ids", None)
outputs = model.generate(**inputs, max_new_tokens=32768, temperature=1.0, top_p=0.95, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Best Practices

1. Reasoning effort: always high. All reported results use high reasoning effort. Pass {"chat_template_kwargs": {"reasoning_effort": "high"}} on every request.
2. Sampling parameters. temperature=1.0, top_p=0.95.
3. Serving. Use the validated SGLang recipe above: BF16, TP=2, FlashAttention-3. Full recipes for every K2-Horizon size, with measured H200 latency and throughput, are in the SGLang cookbook and the vLLM recipe.
4. Parsers. Enable the k2_horizon reasoning parser for chat, and add the k2_horizon tool-call parser for agent use. Leave both off for plain completion-style generation.

Citation

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