Summary
The Institute of Foundation Models — the AI research arm of MBZUAI, a university in Abu Dhabi with satellite labs in Paris and Silicon Valley — released K2 Horizon, a family of six fully open-source models ranging from 0.9 billion to 375 billion parameters. All releases include Apache 2.0 weights, training data, checkpoints, and code. Fahd Mirza tests the three smallest sizes (0.9B, 7B, and 32B) locally on a system with 96GB of VRAM, serving them via vLLM.
Each model size occupies a different niche according to the benchmarks: the 0.9B leads on AM 2026 math and tool calling (BFCL), making it viable for edge and on-device use; the 7B excels on SWE-bench and Terminal Bench for local coding workstations; the 32B pulls ahead on GPQA Diamond graduate-level science reasoning (82 vs 27 for smaller sizes). The 7B and 32B both support 512K context windows. Mirza notes particular interest in the 3.7B, which he calls the “sweet spot” and plans to cover separately.
The honest hands-on verdict is mixed. The 32B took over 30 minutes to produce a rotating kebab skewer animation in HTML and delivered buggy output that didn’t run without manual fixes. Multilingual testing exposed serious gaps in low-resource languages — Bengali, Tamil, Punjabi, and Southeast Asian languages produced garbled mixed-script output. Tool-call parsing also needs work. Mirza frames these as expected limitations for a “stage one” checkpoint, but questions why non-final weights were publicly released.
📺 Source: Fahd Mirza · Published September 04, 2026
🏷️ Format: Benchmark Test
