Tencent Open-Sources Hunyuan Hy4 Preview: 770B MoE with 1M Context

Tencent's Hunyuan team released and open-sourced Hy4 Preview on August 27-28: a 770B-parameter MoE model with 49B active parameters, doubling its predecessor Hy3 (295B/21B/256k), supporting a 1M-token context window, and fully open-sourced under Apache 2.0 on HuggingFace and GitHub with FP8 weights, a blog post, and an arXiv paper. It is positioned as an open frontier model for productivity, priced at 6 RMB per million input tokens, 18 RMB per million output tokens and 0.3 RMB for cache, and is available on WorkBuddy with a free two-week trial.

Confirmed

• Scale: 770B total / 49B active parameters; 1M context; Apache 2.0 with FP8 weights and arXiv paper; two-week free trial on WorkBuddy
• Architecture (per @aigclink citing the GitHub repo): 78 layers, first layer dense FFN, remaining 77 MoE layers each with 256 routed experts plus 1 shared expert, MTP layers and Gated DSA attention
• Official benchmarks show performance roughly on par with GLM-5.3 across 40+ benchmarks; internal blind tests slightly ahead of GLM-5.3 and Kimi K series; per @eyishazyer, it beats DeepSeek V4 Pro on Terminal Bench 2.1
• Per @SimplyAnnisa, the model is more than twice the size of Hy3 released two months earlier; per @aigclink, it introduces a recursive self-improvement loop, the first time the model participates in its own training, targeting real workloads such as coding, games, finance, and scientific simulation
• Official demos show one-sentence generation of high-fidelity 3D scenes; per @iamfakhrealam, Tencent used a "model and product shaped together" strategy with multi-domain expert data
• The official tagline "Use it. Tell us what breaks" encourages integrating the model into agent workflows
• @AlchainHust notes that with this release, Tencent, Alibaba, and ByteDance all reach roughly Opus-4.8-level usability in coding and agentic capability, still months behind SOTA, shifting focus to harness, skill, and agent engineering

Unconfirmed

• Benchmark claims rest mainly on official data and internal blind tests without broad third-party validation; users such as @davidthesong are still asking about real-world experience; some posts misstate total parameters as 77B (official figure is 770B), and context-length reports vary between 100K and 1M (official: 1M)

Why it matters

• A fully open 770B model with 1M context, FP8 weights, and aggressive pricing lowers the barrier to frontier capabilities and directly pressures domestic competitors like GLM, Kimi, and DeepSeek; if the recursive self-improvement mechanism holds up, it offers a new template for training paradigms; @AlchainHust's assessment signals that competition among top Chinese labs is shifting from raw model capability to engineering and deployment.

2026-08-27 ~ 2026-08-29 · 28 related posts

• [source] Tencent Releases Hy4-Preview: New MoE-Based Conversational Model — tencent · 2026-08-27
• Tencent releases Hy4preview: 770B open-source model with 1M context — 腾讯混元 · 2026-08-28
• Tencent open-sources Hy4 Preview: 770B params, 1M context, Apache 2.0 — xeophon · 2026-08-28
• Tencent open-sources Hy4 Preview: 770B params, 1M context, on par with GLM-5.3 — xeophon · 2026-08-28
• [source] Tencent Hunyuan Releases Hy4 Preview: 770B Params, 1M Context — TencentHunyuan · 2026-08-28
• Tencent releases Hy4preview: 770B params, 1M context, and Unreal Engine game generation — 赛博禅心 · 2026-08-28
• Tencent Releases Hunyuan Hy4: 770B Params, CodingAgent Enters First Tier — APPSO · 2026-08-28
• Tencent open-sources Hy4 preview: 770B MoE with 1M context that helped train itself — aigclink · 2026-08-28
• Hy4 preview repo goes live: 256 experts per MoE layer, native MTP head, Gated DSA attention — aigclink · 2026-08-28
• Tencent open-sources Hunyuan Hy4-preview: 770B params, 1M context, ~#3 among open models — 机器之心 · 2026-08-28
• [source] Tencent opens Hy4 preview weights: 770B MoE, 49B active, 1M context — Snoo26837 · 2026-08-28
• Tencent Open-Sources 770B MoE Hunyuan Model, Hits #5 on Arena Coding — 机器之心 · 2026-08-28
• Tencent Open-Sources Hy4-preview: 770B Params, 1M Context, Strong Coding & Research Gains — 智东西 · 2026-08-28
• Tencent Releases Hy4 Preview: 770B MoE with 1M Context — ChrisGPT · 2026-08-28
• Tencent releases Hy4: 770B MoE open-weight model with 1M context — haider1 · 2026-08-28
• Tencent Releases Hy4: 770B Params, 1M+ Context, Aggressive Pricing — eyishazyer · 2026-08-28
• Tencent Releases Hy4 Preview: 770B Total Params, 1M Context, Open Weights — bdsqlsz · 2026-08-28
• Tencent's Hunyuan Hy4 joins Alibaba & ByteDance in crossing coding usability threshold — AlchainHust · 2026-08-28
• Tencent Hunyuan Releases 770B Open Source Model Hy4 — aftahi_ai · 2026-08-28
• Tencent reportedly releases Hy4 preview: 770B params, open source — SimplyAnnisa · 2026-08-28
• Tencent Hunyuan Hy4 launches: 770B params, 1M context, beats Kimi in tests — iamfakhrealam · 2026-08-28
• Tencent Hunyuan rolls out Hy4 Preview, its latest generation model — YouJiacheng · 2026-08-28
• Tencent releases Hunyuan Hy4 preview: 770B MoE with 1M context — ShunyuYao12 · 2026-08-28
• Tencent releases Hy4 Preview model, performance comparable to GLM 5.3 — davidthesong · 2026-08-29
• Tencent open-sources Hy4 preview: 770B params, 49B active, 1M context under Apache 2.0 — TencentHunyuan · 2026-08-29

3 near-duplicate retellings: Scobleizer · AlchainHust · TencentHunyuan