K2 Horizon 7B
K2 Horizon 7B by the MBZUAI Institute of Foundation Models is a fully open dense 7B reasoning model with 512K context, reasoning-effort control, XML tool
Model Overview
Capabilities, design details, and architectural traits
K2 Horizon 7B - the fully open 7B member of the K2 Horizon fleet
K2 Horizon 7B is a dense 7B decoder-only model from MBZUAI's Institute of Foundation Models. It is the medium dense member of the six-model K2 Horizon fleet, which spans 0.9B to 375B-A23B under a shared architecture, vocabulary, and training methodology. Its defining trait is full-lifecycle openness: weights, training data or recipes, training code, and intermediate checkpoints are all public, so capability changes can be studied across training rather than at a single checkpoint.
| Trait | Detail |
|---|---|
| Fleet position | Dense 7B member of the six-model K2 Horizon fleet (0.9B, 3.7B, 7B, 32B, 36B-A4B, 375B-A23B) |
| Context window | Native 524,288-token (512K) context from the midtraining stages onward |
| Openness | Apache 2.0 weights and code, plus released training data and recipe, training code, and intermediate checkpoints |
| Reasoning control | Per-request reasoning_effort via chat_template_kwargs; high is used for all reported results, with thinking returned in reasoning_content |
| Tool calling | Three selectable formats: json, xml, and xml_typed, with xml as the default |
| Recommended settings | temperature=1.0, top_p=0.95, and at least 32,768 output tokens so reasoning is never cut off |
Serving and interfaces
The model ships with day-zero recipes for vLLM and SGLang, both using dedicated k2_horizon reasoning and tool-call parsers. GGUF builds are provided for llama.cpp, requiring a build with K2 Horizon architecture support. It also runs in Transformers with trust_remote_code.
Why the openness matters
IFM positions K2 Horizon as a fleet where the full development process through reasoning and agentic post-training is exposed. For the 7B this is complete today: researchers can reproduce the methods behind its reasoning, tool use, and agentic behavior, or adapt them to new tools and domains instead of treating the final checkpoint as an opaque starting point.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | MBZUAI Institute of Foundation Models | Anthropic | Anthropic |
| Release Date | September 3, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 524K | 1M Best Context Window | 1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | Free Best Input Pricing | $4 | $2 |
| Output Pricing | Free Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | text | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 20.6 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | 38.6 | - | - |
| Agentic Index | 26.1 | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from MBZUAI Institute of Foundation Models
Other models by MBZUAI Institute of Foundation Models
K2 Horizon 375B A23B
K2 Horizon MoVA 36B A4B
K2 Horizon 3.7B
K2 Horizon 0.9B
Top AI Models
Leading alternatives by intelligence score
