K2-Horizon-7B

K2-Horizon-7B is a 7B-parameter dense decoder-only model developed by the Institute of Foundation Models (IFM). Positioned as the medium dense engine in IFM's six-model K2 Horizon family, it targets practitioners looking for high-performance agentic autonomy, software engineering capabilities, and deep mathematical reasoning on local hardware. Unlike many open-weight releases that keep post-training pipelines proprietary, IFM provides full transparency under an Apache 2.0 license, publishing the complete training pipeline, data recipes, intermediate checkpoints, and serving configurations.

Despite sitting in the standard 7B weight class, K2-Horizon-7B redefines what developers can expect from a local AI model with 7B parameters in 2026. Out of the box, it features an architectural context window of 524,288 tokens (512K) established during midtraining. Benchmark evaluations place it ahead of several larger sub-15B competitors, showing exceptional proficiency on complex coding pipelines, long-horizon tool execution, and competition-level mathematics.

For engineers evaluating self-hosted models, K2-Horizon-7B fills the gap between lightweight utility models and heavy MoE networks. It provides the low inference latency and straightforward deployment characteristics of a dense 7B architecture while delivering reasoning scores that directly challenge models nearly twice its size.

K2-Horizon-7B uses a refined dense autoregressive transformer architecture. By opting for a standard dense design rather than a Mixture of Experts (MoE), IFM ensures predictable memory bandwidth utilization and consistent compute throughput across all token generation stages.

The defining technical attribute of the model is its native 524,288-token context window. Rather than relying solely on post-hoc RoPE scaling at inference time, IFM integrated this extended context across its midtraining stages. This minimizes attention degradation and needle-in-a-haystack retrieval loss across deep prompt depths.

However, running full 512K context prompts locally requires careful consideration of the Key-Value (KV) cache:

• At 16-bit precision, an uncompressed KV cache for 524,288 tokens quickly surpasses the memory footprint of the model weights themselves.
• Serving engines like vLLM and SGLang natively support K2-Horizon-7B with optimized PagedAttention, FlashAttention-3 kernels, and FP8/FP4 KV cache quantization to manage this context envelope efficiently.
• Serving configurations ship with dedicated parser integrations for handling multi-turn reasoning chains and structured tool calls natively.

IFM also validated diffusion-based adapters (such as K2-Horizon-7B-Uno) for speculative decoding pipelines, allowing developers to compress decode latency during memory-bound local execution without destabilizing output quality.

K2-Horizon-7B is tuned specifically for agentic environments, technical problem-solving, and code synthesis. Rather than focusing purely on open-ended conversational chat, the model emphasizes structured deterministic execution and rigorous chain-of-thought processing.

The model sets a high bar on software engineering benchmarks, scoring 70.6% on SWE-bench Verified and 39.1% on Terminal-Bench 2.1. This makes it an exceptional engine for autonomous developer agents. Concrete applications include:

• Repository-level refactoring and bug patching across multi-file codebases ingested directly into its 512K context.
• Interactive terminal control, shell command execution, and CI/CD diagnostic remediation.
• Translating complex technical specifications into functional unit tests and implementation code.

With scores of 25.8% on tau3-Banking and 59.0% on BrowseComp, K2-Horizon-7B behaves reliably inside agentic harnesses. It respects strict JSON schemas, outputs well-formed API payloads, and maintains task state across dozens of sequential tool interactions. It excels in local enterprise workflows requiring autonomous web search synthesis, database query construction, and automated ticket triage.

On competition-level mathematics, K2-Horizon-7B recorded a 73.3% score on HMMT Feb 2026, alongside an 18.6% on the Humanity's Last Exam (HLE) scientific benchmark and 31.6% on SciCode. These results make it an ideal local copilot for scientific computing, algorithmic problem solving, and formal logic verification where cloud data privacy cannot be compromised.

To run K2-Horizon-7B locally, practitioners need to plan around two variables: model weight storage and KV cache allocation for their target context size.

The weights themselves fit easily on standard consumer hardware, but context expansion determines your overall VRAM ceiling:

• FP16 / BF16 (Unquantized): Requires ~15 GB VRAM for weights alone. Recommended: 24 GB GPU (RTX 3090, RTX 4090) to accommodate an 8K-32K context buffer.
• Q8_0: Requires ~8.5 GB VRAM. Runs comfortably on 12 GB cards (RTX 4070, RTX 3060 12GB) for standard prompt lengths.
• Q4_K_M (Recommended): Requires ~5.2 GB VRAM for weights. Leaves generous headroom on 8 GB, 12 GB, and 16 GB GPUs for context expansion and local runtime overhead.

Note on 512K Context: Utilizing the entire 524,288 context window requires substantial VRAM for the KV cache. To run the full 512K context locally, you will need a dual-GPU setup (such as 2x RTX 3090/4090 24GB with FP8 KV cache) or a unified memory system such as a Mac Studio with 64 GB to 128 GB of RAM.

• Best GPU for K2-Horizon-7B: An Nvidia GeForce RTX 4090 (24 GB) or RTX 3090 (24 GB) offers the ideal balance, providing enough memory to run the model at full precision or run a Q4/Q8 quantization with deep context windows.
• Mid-Tier Consumer GPUs: An Nvidia RTX 4070 (12 GB) or RTX 4060 Ti (16 GB) runs 4-bit and 8-bit quantizations with zero offloading bottlenecks up to 32K context.
• Apple Silicon: Macs equipped with M3/M4 Pro or M4 Max chips (36 GB to 64 GB unified memory) handle K2-Horizon-7B with large context allocations effortlessly.

For everyday local deployment, Q4_K_M is the best quantization for K2-Horizon-7B. It preserves over 98% of the baseline reasoning capability on SWE-bench and math evaluations while dropping weight overhead down to roughly 5.2 GB. If your workload involves dense scientific logic or competitive math, step up to Q8_0 or FP8 if VRAM permits.

K2-Horizon-7B tokens per second vary by runtime framework and hardware:

• RTX 4090 (vLLM / SGLang, Q4_K_M): Expect 110-140 tokens per second during decode phases with sub-second time-to-first-token on short prompts.
• Apple Silicon (M4 Max, llama.cpp): Expect 75-95 tokens per second at Q4_K_M.
• Mid-Range PC (RTX 4070 12GB, Ollama): Typically delivers 60-80 tokens per second.

The quickest way to get started with the model locally is via Ollama:

1ollama run ifm/k2-horizon-7b

For agentic tool use and maximum throughput, deploy the model through vLLM:

1vllm serve IFM/K2-Horizon-7B \
2  --max-model-len 32768 \
3  --gpu-memory-utilization 0.95 \
4  --enable-reasoning \
5  --tool-call-parser hermes

Evaluating K2-Horizon-7B performance against common alternatives in the 8B to 12B range demonstrates how aggressively IFM tuned this architecture:

Qwen3.5-9B has been a standard benchmark in this tier, but K2-Horizon-7B outperforms it decisively across critical engineering vectors:

• Coding: K2-Horizon-7B scores 70.6% on SWE-bench Verified versus Qwen3.5-9B at 50.8%.
• Mathematics: K2-Horizon-7B reaches 73.3% on HMMT Feb 2026 compared to 65.7% for Qwen3.5-9B.
• Agentic Actions: On Terminal-Bench 2.1, K2-Horizon-7B scores 39.1% against Qwen3.5-9B's 29.2%.
• Trade-off: Qwen3.5-9B has wider community finetune variations, but K2-Horizon-7B delivers stronger raw execution out of the box with a smaller parameter footprint (7B vs 9B).

Google's Gemma 4-12B requires roughly 40% more VRAM to run at equivalent precision levels, yet falls behind K2-Horizon-7B on key technical tasks:

• SWE-bench Verified: K2-Horizon-7B (70.6%) more than doubles Gemma 4-12B (30.6%).
• Competition Math: K2-Horizon-7B leads 73.3% to 63.1% on HMMT.
• Trade-off: Gemma 4-12B benefits from broad ecosystem optimization in mobile toolkits, but K2-Horizon-7B is much lighter on consumer hardware, faster to run locally, and significantly better suited for agentic tool use.

For developers seeking an unencumbered, Apache 2.0 licensed model that runs comfortably on a single consumer GPU while outperforming 9B-12B alternatives on code and math, K2-Horizon-7B is one of the strongest 7B deployments currently available.