{"slug": "tencent-open-sources-hy3-295b-moe-model", "title": "Tencent open-sources Hy3 295B MoE model", "summary": "Tencent released Hy3, an Apache-2.0 open-weight Mixture-of-Experts model with 295B total parameters, 21B active parameters per token, and a 256K context window. The release includes official artifacts on Hugging Face and GitHub, with FP8 variants and serving recipes for vLLM and SGLang. This move signals major Chinese labs packaging frontier-scale MoE systems for practical testing, though independent evaluations are needed to validate performance claims.", "body_md": "# Tencent open-sources Hy3 295B MoE model\n\nTencent released **Hy3**, an Apache-2.0 open-weight Mixture-of-Experts model with **295B total parameters**, **21B active parameters** per token and a **256K context window**. The official Hugging Face card says Hy3 follows the April preview, adds stronger post-training, and ships current weights plus FP8 variants for teams using vLLM or SGLang. For practitioners, the practical question is not only benchmark ranking but serving economics: MoE routing, MTP speculative decoding and a smaller active-parameter path could make frontier-style reasoning experiments more reachable, while the full model still needs large multi-GPU infrastructure. Treat Tencent's benchmark and reliability claims as promising vendor evidence until independent latency, tool-use and production evaluations catch up.\n\nHy3 matters less as another large-model announcement than as an operational signal: major Chinese labs are packaging frontier-scale MoE systems with official artifacts, permissive licensing, and deployment recipes that engineering teams can actually test. The release gives practitioners a concrete model to benchmark against GLM, DeepSeek, Qwen and other open-weight systems when reasoning, long context, and agent reliability matter more than chatbot polish.\n\n### What happened\n\nTencent's official Hugging Face model card lists Hy3 as a 295B-parameter Mixture-of-Experts model with 21B active parameters, 3.8B MTP layer parameters, 192 experts with top-8 routing, BF16 weights, and a 256K context window. The current Hy3 card says it follows the April Hy3 Preview release, moves the model under Apache License 2.0, and publishes Hy3 plus Hy3-FP8 weights across Hugging Face, ModelScope, GitCode and CNB.\n\n### Technical context\n\nThe active-parameter path is the key deployment detail. A 295B MoE can still be expensive to host, but activating 21B parameters per token changes the cost profile compared with dense models of similar headline size. Tencent also documents vLLM and SGLang serving paths with MTP enabled, which makes the release more useful to teams that want to measure real throughput instead of only reading benchmark tables.\n\n### For practitioners\n\nUse the official model card and GitHub repository as the source of truth, then validate the claims on your own workload. Tencent reports gains in agent, coding, long-context and product feedback tests, but the most important production checks are latency under realistic context lengths, tool-call stability, memory footprint, quantization behavior, and whether Apache-2.0 licensing fits internal open-weight policy.\n\n### What to watch\n\nIndependent evaluations should determine whether Hy3's self-reported reliability improvements hold across non-Tencent agent scaffolds and enterprise tasks. If the FP8 and serving recipes mature quickly, Hy3 could become a credible open-weight comparison point for teams that need long-context reasoning without relying solely on closed frontier APIs.\n\n## Key Points\n\n- 1Tencent's current Hy3 release moves the model from preview licensing to Apache-2.0 weights with official Hugging Face and GitHub artifacts.\n- 2The 295B MoE design activates 21B parameters per token, making serving cost the practical test for adopters.\n- 3Model-card benchmark claims are promising, but production teams still need independent latency, reliability, and tool-use evaluations.\n\n## Scoring Rationale\n\nTencent's Apache-2.0 Hy3 release is a major open-weight LLM event because it combines a 295B MoE footprint, 21B active parameter path and 256K context with official model artifacts. The score stays below industry-shaking because benchmark, reliability and serving-cost claims still need broad independent validation.\n\n## Sources\n\nPublic references used for this report.\n\nPractice interview problems based on real data\n\n1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.\n\n[Try 250 free problems](/problems)", "url": "https://wpnews.pro/news/tencent-open-sources-hy3-295b-moe-model", "canonical_source": "https://letsdatascience.com/news/tencent-open-sources-hy3-295b-moe-model-c3d05258", "published_at": "2026-07-07 00:42:01+00:00", "updated_at": "2026-07-07 02:35:16.047860+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "ai-infrastructure"], "entities": ["Tencent", "Hy3", "Hugging Face", "GitHub", "ModelScope", "vLLM", "SGLang", "Apache License 2.0"], "alternates": {"html": "https://wpnews.pro/news/tencent-open-sources-hy3-295b-moe-model", "markdown": "https://wpnews.pro/news/tencent-open-sources-hy3-295b-moe-model.md", "text": "https://wpnews.pro/news/tencent-open-sources-hy3-295b-moe-model.txt", "jsonld": "https://wpnews.pro/news/tencent-open-sources-hy3-295b-moe-model.jsonld"}}