MBZUAI K2 Horizon Claims #1 Speed on Artificial Analysis at 1,067 Tokens per Second
K2-Horizon-375B-A23B is a sparse Mixture-of-Experts model from MBZUAI’s Institute of Foundation Models (IFM) in Abu Dhabi, storing 375B parameters and routing 23B per token. It launched on September 3, 2026, as one of the most open large models released to date — weights, training data, and training code all public. With Artificial Analysis now benchmarking on a 10,000-input-token workload that better reflects production agentic sessions, K2 Horizon has climbed to the top of the output speed table at 1,067 tokens per second through IFM’s free API.
Speed Position on Artificial Analysis
Artificial Analysis tracks output throughput continuously across live API providers. Under its updated benchmark workload, K2 Horizon holds the #1 spot among 78 measured models, co-leading with Celeris-1. The IFM API delivers a median time-to-first-token of 2.28 seconds and completes a 500-token response in 4.62 seconds. Pricing is $0.00 per million tokens.
That combination — first in throughput, zero cost — is not currently matched by any other production model in the rankings.
Reasoning Benchmarks
IFM evaluated K2 Horizon in BF16 and NVFP4 quantization at high reasoning effort:
| Benchmark | BF16 | NVFP4 |
|---|---|---|
| GPQA Diamond | 85.80% | 85.45% |
| AIME 26 (avg@32) | 94.38% | 93.65% |
| MMLU-Pro | 84.22% | 83.98% |
| MBPP | 97.00% | 96.60% |
| GSM8K | 96.06% | 95.53% |
| BBH (3-shot) | 94.73% | 94.26% |
| IFEval (Prompt) | 90.02% | 88.72% |
GPQA Diamond at 85.8% and AIME 26 at 94.4% (averaged across 32 samples) put K2 Horizon’s reasoning in competitive territory with models that run several times as many active parameters. The NVFP4 quantized variant loses under 1 point across all tasks. It requires NVIDIA Blackwell B-series GPUs with native NVFP4 support.
On the Artificial Analysis Intelligence Index, K2 Horizon scores 31, placing it #18 out of 121 tracked models and well above the open-weight median of 19 for comparable sizes. The frontier cluster sits at 70-plus. K2 Horizon is not a frontier intelligence model by that measure, but for throughput-intensive workloads where reasoning depth is sufficient, the cost-performance gap over frontier alternatives is substantial.
Architecture and Openness
The model uses a native 524,288-token context window built in from midtraining. IFM released intermediate training checkpoints at regular steps alongside the final weights, enabling study of how capabilities shift across training — an unusual level of transparency for a model at this parameter scale.
Serving is validated on 8x H200 nodes via vLLM and SGLang. The custom k2_horizon reasoning and tool-call parsers must be enabled for agentic use. reasoning_effort="high" is required to reproduce published benchmark numbers; medium and low effort settings are not recommended for production.
The quantized variant is at IFM/K2-Horizon-375B-A23B-NVFP4 on Hugging Face. Only routed-expert linear layers are quantized to NVFP4; attention, shared experts, routers, and the first three dense layers stay in BF16, which explains the sub-1-point degradation across tasks.