Skip to content
DigIntoAI/Model intelligence
Model intelligence feed

Model intelligence.

Follow the changes. Understand what comes next.

Editor’s pickTencent
Tencent/Tencent Hy4

Tencent Opens Flagship Hy4 Weights (770B Total / 49B Active): Variant Selection Guide and Unity, Unreal, and Godot Adaptation

On November 20, 2026, Tencent opened weights for its flagship multimodal MoE model Hy4 (originally announced November 4, distinct from the lighter 220B Hy4 Flash line) across Hugging Face, ModelScope, and GitCode in three variants: tencent/Hy4, tencent/Hy4-FP8, and tencent/Hy4-base. Featuring a 770B language backbone with 49B activated parameters per token and 1M context, Hy4 adds dedicated adaptation for Unity, Unreal Engine, and Godot.

Latest stories09

Newest published
01

Tencent Opens Flagship Hy4 Weights (770B Total / 49B Active): Variant Selection Guide and Unity, Unreal, and Godot Adaptation

On November 20, 2026, Tencent opened weights for its flagship multimodal MoE model Hy4 (originally announced November 4, distinct from the lighter 220B Hy4 Flash line) across Hugging Face, ModelScope, and GitCode in three variants: tencent/Hy4, tencent/Hy4-FP8, and tencent/Hy4-base. Featuring a 770B language backbone with 49B activated parameters per token and 1M context, Hy4 adds dedicated adaptation for Unity, Unreal Engine, and Godot.

Tencent Hy4Read story
02

Tencent Hy4 flash vs. Hy4 Selection & Pricing Guide: Workload Routing, 50% Discount Deadlines & Preview Retirement

On November 16, 2026, Tencent launched Hy4 flash for high-speed, cost-efficient everyday workflows alongside its flagship Hy4 model. This guide compares official workload recommendations, TokenHub RMB catalog rates (¥0.50/¥2.00 for Hy4 flash vs. ¥6.00/¥18.00 for Hy4 per 1M tokens), separate 50% discount deadlines (Nov 30 vs. Dec 31 Beijing Time), research-page USD reference rates, and the December 7, 2026 retirement of hy4-preview.

hy4-flash; hy4Read story
03

Step 5 Preview Cache Hit Price Doubles Nov 6—Why Your Total Bill Might Not

Tencent Cloud's TokenHub & CloudBase notice raises Step 5 Preview (step-5-preview) cache-hit input from 2026-11-06 00:00:00 Asia/Shanghai: Guangzhou 0.35→0.70, Singapore 0.36→0.72 RMB/M tokens. Miss input and output stay unchanged (numbers not listed). As of 2026-11-04, still a preview—not yet live. A doubled cache-hit line ≠ a doubled total bill; impact tracks hit share. TokenHub/CloudBase only, not a StepFun all-channel hike. DigIntoAI has not checked bills or hit rates—figures follow the Tencent Cloud notice.

step-5-preview/TokenHub/CloudBase/缓存命中输入价Read story
04

Tencent Hy4 Hits TokenHub and OpenRouter—Vision In, 50% Off Through Year-End

TokenHub listed formal Hy4 (hy4) 2026-11-04 12:13:30, adding image understanding vs preview; 1M / 960k / 64k; list 6 / 18 / 0.3 RMB/M with 50% through 2026-12-31 23:59:59 Beijing time (~3 / 9 estimate). OpenRouter tencent/hy4 2026-11-04 12:20:31 Asia/Shanghai: ~$0.417 / $1.251/M, 50% off, text+image→text. Hy4 preview (hy4-preview) leaves TokenHub/ADP/CloudBase from 2026-12-07 00:00:00 Beijing time; conditional auto-switch to hy4, else call errors.

Hy4/hy4/Hy4 preview/TokenHub/OpenRouter tencent/hy4Read story
05

What You Still Need to Reproduce Tencent Hyra HyInv and HyPLM-Flash

At 16:50 Asia/Shanghai on 2026-11-02 (08:50 UTC), Tencent-Hunyuan/Hyra-results merged HyInv and HyPLM-Flash under AI4Science, and published formal HyPLM-Flash FP32 weights as Release hyplm-flash-weights (prerelease: false). Open inference plus weights is not a full training reproduce. Environment, precision paths, and reproducibility pins below follow the commit, Release, and READMEs—we have not measured speed or accuracy ourselves.

HyInv/HyPLM-Flash/Hyra 2.0Read story
06

Tencent Opens Hy-SimulMT — Download Isn't Instant Interpreting

On October 26, 2026, Tencent Hunyuan released Hy-SimulMT for bidirectional Chinese–English real-time simultaneous translation with real-time voice cloning. Code and weights are public under Apache-2.0; the main checkpoint is Hy-SimulMT-30B-A3B (~30.4B). RealSI reports StreamLAAL of about 3 seconds—a streaming lag metric, not time-to-first-speech. Speech output and cloning need a configured TTS service. A llama.cpp PR (#31198) is still open and unmerged; Hugging Face lists no inference provider. Open weights are not “plug in and interpret.”

Hy-SimulMTRead story
07

Hunyuan Opens Hy-Nomas: Interactive Video World Model on a 4090

Tencent Hunyuan has open-sourced Hy-Nomas, a free-form interactive video world model. Users can steer successive video segments with text instructions and camera trajectories, with a persistent world memory across turns. Weights and code are under Apache-2.0 on GitHub and Hugging Face, with an official website and interactive demo. A single RTX 4090 path exists as a low-memory configuration—it is not the same as the standard eight-GPU setup.

Hy-NomasRead story
08

Tencent Yuanbao Search WSA Pricing (October 2026)

Tencent Cloud’s Yuanbao Search (WSA) exposes API and MCP access with official pay-as-you-go tiers at ¥14 / ¥20 / ¥26 per 1,000 calls for Speed / Standard / Flagship. Flagship is enterprise-only on the billing docs. Public docs do not hard-number free monthly quotas; media figures must be checked in the console activity zone.

Yuanbao Search / WSA / SearchPro / MCPRead story
09

AI coding plans compared: prices, usage limits and buying in Japan

Compare coding subscriptions by monthly price, included usage and extra charges. The November 20 review covers Qwen, Copilot, Cursor, Codex and Claude Code, with Japanese billing notes and the October edition.

Codex / Claude Code / Antigravity / Cursor / GitHub Copilot / GLM / Kimi / Qwen / MiniMaxRead story