K2 Horizon 375B A23B

K2-Horizon-375B-A23B

Quick answer: K2-Horizon-375B-A23B is IFM's flagship open-weight Mixture-of-Experts model — 375B total parameters, 23B active per token, with a native 512K-token context window. On agentic tool-use, terminal, and long-horizon workflow benchmarks it matches or beats open-weight MoE models up to 2.6x its size and is competitive with closed frontier models. Open weights, Apache 2.0 license.

At a Glance

Where K2-Horizon-375B-A23B leads

• Strong agentic tool use and terminal-use scores relative to its 23B active-parameter footprint
• 512K native context window from midtraining onward
• Fully open: weights, and (upcoming) training data/recipe/code

Where it lags

• Trails closed frontier models (GPT-5.6, Claude Sonnet 5) on GDPVal-AA Elo and BrowseComp
• CritPt (frontier physics reasoning) is notably weaker than most peers in this comparison set

Best for: teams wanting an open-weight model with frontier-adjacent agentic tool-use and terminal-use capability at a fraction of the active-parameter cost of larger MoE peers.

What K2-Horizon-375B-A23B Is

K2-Horizon-375B-A23B is the largest model in IFM's K2-Horizon family, a sparse Mixture-of-Experts architecture storing 375B parameters while activating only 23B per token. It is built for long-horizon agentic workflows — tool use, terminal interaction, and multi-step task completion — rather than pure static-benchmark chasing. The model ships with the final training checkpoint; intermediate checkpoints, training data, and training code are planned for future release, an unusually high transparency bar for a model at this scale.

Specifications

Pricing

Open weights: free to download and self-host; cost is hosting/inference only. Validated serving recipes exist for vLLM and SGLang (8x H200, tensor-parallel 8).

Context Window

K2-Horizon-375B-A23B has a 524,288-token (512K) context window — roughly 1,300+ pages of text in a single request. This is native from the midtraining stages onward, supporting long-document and long-horizon-agent workloads without external context-extension tricks.

Public Benchmark Scores

AA-Omniscience Accuracy (23.0) and Non-Hallucination (74.7) were also reported but use a different metric structure than the platform's combined AA-Omniscience Index — pending mapping decision, not yet added. SWE-Atlas-QnA (48.4, strict) reported but pending exact-benchmark mapping confirmation. Scores above are reported by IFM and shown for context; not Benchgen measurements. GDPVal-AA is the Elo rating; other scores are % unless noted.

Use K2-Horizon-375B-A23B via API

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,
    extra_body={"chat_template_kwargs": {"reasoning_effort": "high"}},
)
print(response.choices[0].message.content)

Frequently Asked Questions

Specs and scores sourced from IFM's official Hugging Face model card; third-party benchmark scores attributed inline. Last updated 2026-09-03.