Robots Atlas>ROBOTS ATLAS
K2 Horizon 32B
AI Modelsโ€บK2 Horizon

K2 Horizon 32B

32Bย ยทย Family: K2 Horizon
Dense 32B model from IFM's K2 Horizon family: 512K context, Apache 2.0, fully open (LLM360), with reasoning and tool use.
โณ Previewโœ“ Public accessโš– Open sourceLLMReasoning model๐Ÿ“ K2 Horizon
Context window
512K
tokens
Parameters
32B (dense, decoder-only)
parameters
Release date
3 September 2026
๐ŸขInstitute of Foundation ModelsProducer
Access:DownloadDeployment:๐Ÿ’ป Localโ˜ Cloud

Overview

K2 Horizon 32B is a dense language model from the K2 Horizon family developed by the Institute of Foundation Models (IFM). Unlike the family's larger sparse Mixture-of-Experts variants, this is a dense decoder-only model with 32B parameters โ€” all activated per token.

The model supports a native 512K context window (524,288 tokens) from the midtraining stages onward. By default it runs with high reasoning effort, supports tool calling, and targets reasoning, coding, and scientific tasks.

K2 Horizon 32B is fully open following the LLM360 principle โ€” IFM plans to release the data, recipe, and training code. Weights are available under Apache 2.0 in safetensors (BF16) format. Serving was validated on 2ร— H200 with tensor parallelism (TP=2) via vLLM and SGLang (with FlashAttention-3). The released checkpoint is a Stage 1 checkpoint; the final checkpoint is pending.

Classification
LLMReasoning model
Family: K2 Horizon
Access & deployment
Download
LocalCloud
Weights: Open source
Key parameters
๐Ÿ“ Context: 512K
๐Ÿงฉ Parameters: 32B (dense, decoder-only)
โœ“ Toolsย ยทย โœ“ Fine-tuning
๐Ÿ“ฅ Input: text

Technical specification

Context window
512K
tokens
Parameters
32B (dense, decoder-only)
parameters
License
Apache 2.0
Hardware requirements
BF16 serving on 2ร— H200 with tensor parallelism (TP=2), e.g. via vLLM or SGLang (with FlashAttention-3). Recommended sampling: temperature 1.0, top_p 0.95.
Features:โœ“ Tool useโœ“ Fine-tuning
Modalities
โฌ‡ Input
text
โฌ† Output
textcode

Capabilities and applications

Native model capabilities
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Tool use
The model's ability to call external functions, APIs and tools during a conversation: calculator, search engine, code editor, database. The model decides when and how to use a tool and interprets its result.
Category: planning
Long context
Support for large context windows โ€” tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Adaptive reasoning effort
The model decides how much 'thinking' to allocate to a given query: simple questions are answered quickly, complex problems receive more inference cycles. A GPT-5.1 feature (both Instant and Thinking) that shortens time on easy tasks and extends it for hard ones.
Category: reasoning

Benchmark results

9 benchmarks
GPQA
accuracy ยท high reasoning effort
82.3%
๐Ÿ“„ IFM
Humanity's Last Exam (HLE)
accuracy ยท high reasoning effort
22.8%
๐Ÿ“„ IFM
Terminal-Bench 2.0
accuracy ยท high reasoning effort
36.6%
๐Ÿ“„ IFM
SciCode
accuracy ยท high reasoning effort
30.2%
๐Ÿ“„ IFM
AA-LCR
accuracy ยท high reasoning effort
65.3%
๐Ÿ“„ IFM
CritPt
accuracy ยท high reasoning effort
1.4%
๐Ÿ“„ IFM
tau3-Banking
accuracy ยท high reasoning effort
22.5%
๐Ÿ“„ IFM
AA-Omniscience Accuracy
accuracy ยท high reasoning effort
16.8%
๐Ÿ“„ IFM
AA-Omniscience Non-Hallucination
accuracy ยท high reasoning effort
58.3%
๐Ÿ“„ IFM

Technical architecture

Sources and related pages