Many teams struggle with large models that need huge GPU memory, produce hallucinations, and break when calling tools across multiple steps. Hy3 solves these issues by being a 295B parameter Mixture-of-Experts model that only activates 21B parameters per token, keeping compute low while still handling 256K-token contexts. Its Apache 2.0 license lets you run it freely, and an FP8 checkpoint further cuts memory usage for cheaper serving.
To avoid hallucinations, Hy3 was trained to answer only when grounded and to flag missing evidence, dropping internal hallucination rates from 12.5% to 5.4%. Tool-call stability was improved so invalid calls that cause loops are rare, and accuracy variance across different agent scaffolds stays within 4%. Multi-turn intent tracking was strengthened with joint SFT and RL, reducing coreference errors from 17.4% to 7.9% and boosting long-dialogue scores from 42.9% to 75.1%.
Deploying Hy3 is straightforward with vLLM or SGLang. Launch the model on eight H20-3e-class GPUs, enable the MTP speculative decoder, and use the OpenAI-compatible API. Control depth of reasoning with the reasoning_effort flag: use no_think for quick answers, low for light reasoning, and high for math, coding, or multi-step tasks. Keep temperature at 0.9 and top_p at 1.0 as recommended. If you lack local hardware, the free tier on OpenRouter provides access until July 21 2026.
Hy3 delivers strong coding (78.0 on SWE-Bench Verified) and reasoning (90.4 on GPQA Diamond) performance while using far less active memory than comparable dense models, making it ideal for agentic workflows, long-contract analysis, financial summarization, and frontend development without the usual cost or reliability headaches.
#AI #LLM #MoE #Agentic #OpenSource #Product