{"rewrite":{"id":"r_51472149cad4efb5d89d427d","clusterId":"c_2426229343df7827ca90f8f6","slug":"tencent-releases-hy4-preview-ai-model-open-and-free","model":"deepseek-v4-flash","headline":"Tencent Releases Hy4 Preview AI Model, Open and Free","summary":"Tencent released the AI model Hy4 preview on August 28, 2026. It is an open model with 770 billion total parameters, 490 billion active parameters, and a 1 million token context length. It surpasses GPT-5.6 Sol on some benchmarks and is free under Apache License 2.0.","whyItMatters":"Tencent positions Hy4 preview as a state-of-the-art open model that beats GPT-5.6 Sol on specific agent and coding tests, while also releasing a quantized 1-bit version that runs on far less VRAM.","webCardHtml":"\u003cp\u003eTencent\u0026#39;s Hy4 preview arrives as an open model with a 770 billion parameter MoE architecture, of which 490 billion are active per token. The model handles up to 1 million tokens of context and is licensed under Apache License 2.0, so it can be downloaded and used freely.\u003c/p\u003e\u003cp\u003eOn benchmarks, Hy4 preview outperforms GPT-5.6 Sol in SWE Atlas Refactoring and SWE-bench Pro, which measure agent and coding performance. Tencent also released quantized versions: the Q4_K_M runs with 435GiB of VRAM, while the STQ1_0 version needs only 214GiB, using a 1.31-bit quantization method that keeps performance loss small relative to the size reduction.\u003c/p\u003e","blueskyPost":"Tencent released Hy4 preview, an open MoE model with 770B total params and 1M token context. It beats GPT-5.6 Sol on SWE-bench Pro and SWE Atlas Refactoring. Also out: a 1-bit quantized version that runs in 214GiB VRAM. Free under Apache 2.0.","twitterPost":"Tencent's Hy4 preview is out: open MoE, 770B params, 1M context. It tops GPT-5.6 Sol on SWE-bench Pro and SWE Atlas Refactoring. Quantized STQ1_0 runs on 214GiB VRAM. Apache 2.0.","threadsPost":null,"newsletterBlurb":"Tencent released Hy4 preview, an open AI model with 770 billion parameters and a 1 million token context. It beats GPT-5.6 Sol on some agent and coding benchmarks. A quantized 1-bit version runs on far less VRAM.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20260831-hy4-preview-tencent/\",\"title\":\"Tencent Releases AI Model 'Hy4 preview', Outperforming GPT-5.6 Sol in Some Tests and Also Releases 1-bit Quantized Version Slimmed Down to 213GB\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4834,"outputTokens":634,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1788143743,"createdAt":"2026-08-31T02:25:07.000Z","publishedAt":"2026-08-31T02:27:33.000Z","updatedAt":"2026-08-31T02:27:33.000Z"},"cluster":{"id":"c_2426229343df7827ca90f8f6","canonicalTitle":"TencentがAIモデル「Hy4 preview」を公開、一部テストでGPT-5.6 Solを上回る＆213GBまで軽量化した1bit量子化版も公開される","representativeArticleId":"a_0c6e806bbcead49f9296fd72","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[\"Hy4 preview\"],\"studios\":[\"Tencent\"],\"people\":[],\"type\":\"announcement\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-08-31T01:44:00.000Z","lastSeenAt":"2026-08-31T01:44:00.000Z","updatedAt":"2026-08-31T02:27:29.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20260831-hy4-preview-tencent/","title":"TencentがAIモデル「Hy4 preview」を公開、一部テストでGPT-5.6 Solを上回る＆213GBまで軽量化した1bit量子化版も公開される"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":["Hy4 preview"],"studios":["Tencent"],"people":[],"type":"announcement","domain":"other","is_roundup":false},"keyFacts":null}