Hy4 preview

The vals.ai model page for Hy4 Preview records a release date of August 28, 2026, credits developer Tencent, and lists a 1M context window with a 64k max output. It confirms the model is open-weight, takes text and file input (no image or video input), and reports token costs of $0.83 in / $2.50 out, which mirror the O On independent evaluation, vals.ai posts a Vals Index accuracy of 55.40% (±1.25) at a cost per test of $1.281 and a latency of 47 minutes 51 seconds, and breaks down per-benchmark ranks such as Code Migration (175/61), CyberBench (46/26), Finance Agent v2 (14/61), EMB (160/58), SRE Bench (7/10), Tax Agent Bench (11/21)

IntuitionLabs provides an in-depth architectural analysis of Tencent Hy4 preview, confirming it as a Mixture-of-Experts model with 770 billion total parameters and 49 billion activated per token, released on August 28, 2026 under the Apache License 2.0. The model's weights are published in two forms, a full-precision H The analysis notes that Tencent's blind comparative evaluation was conducted internally, with detailed methodological limitations that suggest benchmark standings should be read as a snapshot of a self-selected internal test suite rather than a settled, third-party-audited ranking. Despite these caveats, the article co

AI Profit Boardroom documents Tencent's open-sourcing of Hy4 preview on August 28, 2026, with 770B total parameters, 49B active per token, and a context window beyond one million tokens, framed as a frontier-scale open-weight model. The article notes a two-week free access period inside WorkBuddy and CodeBuddy alongsid The piece walks through benchmark gains and enterprise fit, calling Hy4 preview the most ambitious Tencent release to date and noting that Tencent itself said the model helped optimize its training pipeline. It also connects Hy4 preview to the wider Chinese open-weight wave and explores where it lands in an AI business

Developers Digest frames Hy4 preview as a 770B-total-parameter, 49B-active MoE model released under Apache 2.0 with a 1M-context text-only design, calling it a roughly 2.6x jump over July's Hy3 (295B/21B active, 256K context) and the biggest open-weights release of August 2026. The article reports Terminal-Bench 85.4 a The piece positions Hy4 preview for software engineering, research and financial analysis, contrasting it with cheaper open-weights alternatives. It serves as a secondary decision-intent read covering architecture, benchmarks, API cost, and self-hosting realities rather than a gateway or reseller promotion, and ties ea

Shattered.io reports that Tencent pushed Hy4 preview as a 770B-parameter MoE model with about 49B active parameters onto Hugging Face on August 28, 2026, alongside a lower-precision Hy4 preview-FP8 variant for smaller-footprint local or edge deployment, with mirrors on ModelScope, GitCode and CNB per TechNode. The arti The piece contextualizes the release against concurrent moves from Alibaba, Moonshot AI and Zhipu in the open-source coding market and discusses what Hy4 preview signals about the AI coding-assistant race. It surfaces the strategic weight Tencent places on productivity workloads and notes the rapid developer uptake wit

Tencent's official release announcement on tencent.com confirms that Hy4 preview was released and open-sourced on August 28, 2026, with 770B total parameters, 49B active parameters and a context window exceeding 1M tokens, targeting coding, office work and scientific research. The model is accessible as open-source wei The announcement highlights that Hy4 preview was co-designed with products like CodeBuddy and WorkBuddy and trained with Tencent experts across software engineering, gaming, finance and security domains. In Tencent's internal blind evaluation with 163 experts on 203 engineering tasks, Hy4 preview averaged 2.99/4.00, na

BenchLM's Hy4 preview page catalogs the model with 29 published benchmark rows, a 1M-token context window, and "Self-hosted" infrastructure cost varying, listing an input median of $0.97. The aggregator ranks Hy4 preview at Coding rank 16 of 135 (89th percentile) and Agentic rank 32 of 104 (70th percentile), with verif The page frames Hy4 preview as a self-hosted open-weights model where infrastructure cost varies rather than quoting a first-party hosted token rate, and notes that time-to-first-token and speed are not measured. It sits alongside time-of-publication (Aug 28, 2026) and serves as a third-party tracker of benchmark cover

FelloAI reports that Tencent released Hy4 preview on August 28, 2026 and open-sourced the weights the same day, making it the largest permissively licensed model shipped, with 770 billion total parameters, 49 billion active per token, and a one million token context window under the Apache 2.0 license. The model is des The article details pricing at $0.834 per million input tokens and $2.501 per million output tokens, with cached input at $0.042, and notes that the official deployment recipe assumes 16 B200s or 8 B300s GPUs. Every published benchmark score originates from Tencent itself, with no independent Artificial Analysis rating