K2-Horizon-375B-A23B-FP8

Training Code - Evaluation Code - Pretraining Data - Midtraining Data

This repository contains an FP8-quantized version of IFM/K2-Horizon-375B-A23B.

Only the routed-expert linear layers are quantized to FP8:

• Weights: static FP8, one scale per 128*128 block.
• Activations: dynamic FP8, one scale per 1*128 group along the input-channel dim.

All other linear layers (attention, shared experts, routers, the first 3 dense layers, and lm_head) are kept in BF16.

The FP8 model performs closely in line with the original BF16 model on our evaluations, while reducing memory footprint and enabling faster inference on FP8-capable hardware.

Serving note: the routed experts' intermediate size (1792) is not splittable into whole 128-wide quantization blocks at the usual tensor-parallel sizes (TP=4, TP=8), so expert parallelism is required.

K2-Horizon-375B-A23B is the flagship of the K2-Horizon family: a sparse Mixture-of-Experts model that stores 375B parameters and runs 23B per token, with a 512K context window. We have released the final checkpoint; intermediate checkpoints, along with the data and the training code, will be released.

K2-Horizon-375B-A23B Highlights

• Frontier-class agentic performance. On agentic tool use, terminal, and long-horizon workflow benchmarks it matches or beats open-weight MoE models up to 2.6× its size and is competitive with closed frontier models (see Benchmark Results).
• 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 models				Closed models
	K2-Horizon-375B-A23B	Nemotron 3 Ultra	Inkling (xhigh)	MiniMax-M3	GLM 5.2 (max)	GPT 5.6 Luna (max)	GPT 5.6 Terra (high)	Claude Sonnet5 (max)
# Params	375B	550B	975B	428B	753B	--	--	--
# Activated params	23B	55B	41B	23B	40B	--	--	--
Architecture	MoE	MoE	MoE	MoE	MoE	Closed	Closed	Closed
Agents
GDPVal-AA Real-world professional tasks (Elo)	1,441	1,162	1,234	1,380	1,498	1,569	1,503	1,584
tau3-Banking Agentic tool use	34.0	14.2	29.1	15.3	34.6	31.1	28.7	37.3
Coding
Terminal-Bench 2.1 Agentic terminal use	70.2	53.9	55.1	65.2	77.9	80.9	75.7	80.5
SciCode Scientific coding	42.7	39.9	46.1	45.4	50.5	52.5	50.1	53.6
Scientific Reasoning
Humanity's Last Exam (without tools) Expert-level reasoning	32.0	28.4	31.9	39.0	41.1	39.5	38.5	41.3
GPQA Diamond Graduate-level science QA	87.3	86.7	87.2	92.9	89.5	91.1	89.6	91.1
CritPt Frontier physics reasoning	8.6	3.1	5.4	3.7	20.9	21.0	22.9	16.9
General
AA-LCR Long-context reasoning	76.0	71.0	73.3	80.3	76.7	78.3	73.3	77.0
AA-Omniscience Accuracy Factual accuracy	23.0	23.0	42.0	17.0	24.0	43.0	45.0	40.0
AA-Omniscience Non-Hallucination Non-hallucination rate	74.7	70.0	32.0	82.0	74.0	7.0	10.0	61.0
Agentic Evaluations
Toolathlon Verified Agentic tool use	65.3	34.3	45.5	53.7	59.9	67.5	64.8	71.6
Automation Bench Public Workflow automation	25.3	8.0	12.8	20.5	26.2	33.5	28.0	34.7
Apex-Agents (pass@1) Long-horizon professional workflows	24.8	9.0	19.0	23.8	26.9	28.6	25.4	31.7
MCPMark MCP tool use	67.7	45.7	51.2	48.8	72.4	66.9	74.0	65.3
BrowseComp Deep web research	72.8	44.4	77.1	83.5	--	83.3	--	84.7
WildClawBench In-the-wild agentic tasks	50.9	34.2	52.3	56.4	55.0	50.4	60.0	--
SWE-Atlas-QnA Repo-level code Q&A (strict)	48.4	--	25.5	42.3	46.4	--	--	--
SWE Bench Pro Software engineering (strict)	42.6	38.7	43.1	43.8	46.7	48.8	--	--

Scores in %, except GDPVal-AA, which is an Elo rating. Bold marks the best score in each row. The first four sections follow the Artificial Analysis Intelligence Index categories. Baseline scores are from Artificial Analysis where available, otherwise from the IFM evaluation harness. SWE-Atlas-QnA and SWE Bench Pro are run without internet access; BrowseComp uses the Discard-all@95k context setting from the DeepSeek-V3.2 technical report; WildClawBench and Apex-Agents use the English text-only subsets.

Quickstart

Serving

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

vllm serve IFM/K2-Horizon-375B-A23B \
  --revision main \
  --model-impl transformers \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --trust-remote-code \
  --dtype bfloat16 \
  --max-model-len 131072 \
  --reasoning-parser k2_horizon \
  --tool-call-parser k2_horizon \
  --enable-auto-tool-choice

SGLang recipe validated on 8× H200 in the SGLang K2 Horizon cookbook:

python3 -m sglang.launch_server \
  --model-path IFM/K2-Horizon-375B-A23B \
  --revision main \
  --tp 8 \
  --ep 8 \
  --dtype bfloat16 \
  --attention-backend fa3 \
  --model-loader-extra-config '{"enable_multithread_load":false}' \
  --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, and at least 32,768 output tokens. 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-375B-A23B",
    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 4.57.6, PyTorch 2.13.0, Safetensors 0.8.0.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "IFM/K2-Horizon-375B-A23B"
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=8 on one 8× H200 node, FlashAttention-3, with multithreaded weight loading disabled. 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/},
}