Hy3
Hy3 by Tencent (2026): 295B total / 21B active MoE, 256K context, Apache 2.0. 40% inference efficiency gain via vLLM/SGLang co-design. Beats larger flagships on agentic coding.
Provider: Tencent · Family: Hunyuan
Context window: 256,000 tokens · Max output: 32,768
Input modalities: text, image · Output: text
About Hy3
Hy3 is Tencent's third-generation Hunyuan (混元) large language model, officially released in July 2026 following a preview launch in April 2026. It is a Mixture-of-Experts (MoE) model with 295 billion total parameters and 21 billion active parameters per forward pass, supporting a context window of up to 256,000 tokens. The model implements a hybrid fast-and-slow thinking architecture, allowing it to dynamically allocate compute between quick responses and extended reasoning chains. Hy3 was developed after Tencent rebuilt its pre-training and reinforcement learning infrastructure starting January 2026, guided by three principles: well-rounded capabilities across reasoning, long-context, instruction-following, and tool use; authentic evaluation beyond standard benchmarks; and model-inference co-design for cost efficiency. The 40% inference efficiency improvement comes from deep co-optimization between the MoE architecture and inference frameworks (vLLM, SGLang), including compute performance gains and advanced quantization algorithms. Since its preview release, Hy3 has been integrated across Tencent's product ecosystem: Yuanbao (AI assistant) added agent functions for complex tasks generating PPT, Word, Excel, PDF, HTML files; ima (knowledge workspace) gained structured reasoning and long-form writing improvements; CodeBuddy/WorkBuddy (developer/office productivity) saw 20% PPT generation success rate increase; Marvis (OS-level agent) enhanced file editing, diagnostics, and multi-agent collaboration. Real-world usage shows Hy3 reliably powers agent workflows up to 495 steps across document processing, data analysis, knowledge retrieval, and MCP toolchain orchestration. Hy3 is open-sourced under the commercially friendly Apache 2.0 license, available on Hugging Face, ModelScope, GitHub, and GitCode from day one. It supports mainstream inference frameworks vLLM and SGLang for direct deployment. Global third-party platforms (OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, Cherry Studio) are progressively integrating Hy3. On Tencent Cloud TokenHub, Hy3 preview pricing started at ~$0.18/M input tokens, $0.06/M cached input, $0.59/M output tokens, with personal plans from ~$4.10/month for agent development platforms. Benchmark highlights (Hy3 preview): tops overall usability and agent capability benchmarks; exceptional STEM reasoning on complex tasks; 40% inference efficiency gain at comparable intelligence density. The model demonstrates strong in-context learning, instruction following, coding, and agent capabilities — particularly excelling in repository-level coding tasks within development environments and search-integrated workflows via MCP (Model Context Protocol).
Pricing
Tencent Cloud TokenHub pricing (Hy3 preview): input ~$0.18/M, cached input ~$0.06/M, output ~$0.59/M tokens. Personal agent platform plans from ~$4.10/month. Local inference free (Apache 2.0) — requires GPU (A100/H100). No batch discount published yet.
Key Features
- Hybrid Fast/Slow Reasoning: Dynamic compute allocation between quick responses (fast mode) and extended chain-of-thought (slow mode) — first open-weight MoE with this architecture.
- 40% Inference Efficiency via Model-Inference Co-Design: MoE architecture co-optimized with vLLM and SGLang yields 40% token-generation speedup at same intelligence density — measured on Artificial Analysis.
- 256K Context Window with High Recall: Supports 256,000 tokens context with verified long-context needle-in-haystack performance — enables whole-repository code analysis and multi-document Q&A.
- Apache 2.0 Open Weights with Full Ecosystem Support: Weights, code, training recipes on Hugging Face, ModelScope, GitHub, GitCode. vLLM/SGLang native. Commercial use, fine-tuning, distillation permitted.
- Native MCP and Agentic Tool Use: RL-trained on MCP toolchains; reliably executes 495-step multi-agent workflows across GitHub, filesystem, web search, code execution.
Pros
- Best intelligence-per-dollar among open-weight flagships — 40% efficiency gain + Apache 2.0 = unbeatable cost curve for self-hosted agentic workloads.
- Hybrid reasoning architecture is genuinely novel — not just 'thinking tokens' but dynamic MoE routing between fast/slow expert paths.
- Proven at massive scale inside Tencent: 20x token growth since preview, 495-step agents, 150+ enterprise integrations via TokenHub.
- Full open-source stack: weights, code, recipes, quantization configs — no 'open weight but closed training' ambiguity.
Cons
- Peak reasoning lags Claude Fable 5 (60), Opus 4.8 (56), GPT-5.5 (55) on Intelligence Index (Hy3: 52) — not the absolute smartest model.
- Chinese-first documentation and community creates friction for Western developers; English resources incomplete.
- Tencent Cloud international onboarding (KYC, Chinese-default console) adds friction vs. OpenRouter/Anthropic/OpenAI APIs.
- Vision capabilities unbenchmarked against dedicated multimodal flagships; no audio/video I/O.
Benchmarks
- math: 72
- mmlu: 88
- mmlu pro: 78
- aime 2025: 75
- arc agi 2: 18
- humaneval: 88
- live bench: 58
- lmarena elo: 1320
- gpqa diamond: 62
- lmarena rank: 12
- aider polyglot: 65
- swe bench verified: 55
- humanitys last exam: 15
- artificial analysis intelligence index: 52
- artificial analysis price blended per m: 0.85
- artificial analysis speed tokens per sec: 95