LONG AI
Home › Blog › Tencent Hunyuan Hy4 GA: 770B, First with

Tencent Hunyuan Hy4 GA: 770B, First with Vision and Computer Use

Category:行业资讯 · 15 min read · 2026-11-05

This column aggregates the latest large-model releases and open-source drops from the past one to two days, cross-checked against official announcements and multiple media outlets. Performance figures distinguish between vendor claims and independently verified third-party results; anything not yet replicated is clearly marked. Illustrations are original.

Mechanical brain with camera and robotic hand operating a computer illustration
The model learns to "see" and "act": Hy4 GA adds vision and computer use for the first time (illustration)

Headline: Tencent Hunyuan Hy4 GA ships, filling the vision and computer-use gaps

What was "imminent" yesterday landed today. Per Tencent and multiple media, Tencent Hunyuan released the next-gen Hy4 on Nov 4, claiming the series has entered the strongest open-model ranks for the first time. Key numbers:

Take: this is the first shot in November's domestic-flagship showdown. Tencent played its hand first, unusually printing its own weak scores into the release. Alibaba's Qwen4 hasn't moved yet. Two things to reconcile next: when the weights actually drop, and what third parties score.

Afternoon open source: Baidu Wenxin 5.5 Flash Base, 280B with only 14B active

Hours after this issue first went out, Baidu released and open-sourced the base model Wenxin 5.5 Flash Base (ERNIE-5.5-Flash-Base) on Nov 5, weights on Hugging Face - a genuine new addition.

Take: today's two releases represent two routes - Tencent stacks scale (770B) to chase "strongest," Baidu squeezes per-token cost (280B activating 14B) to chase "cheapest." For agent builders the latter is more practical: long task chains spend on repeated calls and long-context cache, not single answers.

Other releases (5)

Model / ProductDateTypeHighlights
Kimi K3.1 Pro full weights (Moonshot)11-4Open weights (MoE, native multimodal agent)The "open before Nov 4" promise delivered on time: full weights on Hugging Face, 2.8T total / 104B active, ~1M context, released as MXFP4 quantized (MXFP8 activations, quantization-aware training); API open at $2/$10 per million. Moonshot explicitly says it won't release weights that push network-attack capability to the frontier (removed some security training data) - a "bounded openness" template.
Apollo World (Baidu Apollo)11-3Physical-AI world model (autonomous driving)First disclosed at Baidu AI Day: models vehicles and pedestrians in a virtual city as decision-capable AI agents, generating camera and LiDAR observations in realtime - an end-to-end "AI teaches AI to drive" loop, already used in Apollo Go's global testing (28 cities, 240M km fully driverless). "100K km/day" virtual road-testing equals ~3 years of a real car. Vendor-reported, unreplicated.
Images 2.6 + GIF generation (OpenAI)11-4/5Image model + productChatGPT upgrades to Images 2.6 with direct GIF generation and editing (speed, per-frame, frame-by-frame), all platforms; two image APIs ship: GPT-Image-2.6 Flare (fast, ~6s low-quality) and Sunburst (quality + multi-turn consistency). Vendor: ChatGPT + Codex + API over 4B images/week (was 3B in September).
FlashProve (Anthropic, open source)11-2/11-4Provable inference (cryptographic)Uses zero-knowledge proofs to give model answers a "mathematical birth certificate": providers pre-commit a public ID for their weights; users verify the answer came from the specified model on a normal CPU in ~17s, without the weights or rerunning. On Llama 4 Maverick (400B), one H100 proves ~110 tokens/s, ~18.78MB per proof; the cost is 70-95x vLLM GPU time, 1024-token context, and it proves source not correctness. Apache 2.0.
OpenAI dual move: Pierre Computer + GPT-5.6 Cyber to gov cloud11-4M&A / gov modelAcquires Pierre Computer Company (folding its dev tools into Codex, undisclosed); GPT-5.6 Cyber enters AWS GovCloud via Bedrock for authorized vuln research and security testing. GPT-6.2 Astra gray-test leaks and rc-ada-* IDs briefly appearing are unconfirmed.

Industry news

The trend in one line

Today's keyword is "delivery," subtitled "efficiency." Hy4 delivered the Wuzhen promise, Kimi delivered its open-source promise, Apple delivered its domestic-AI timeline - while Baidu's 280B-activating-14B base reminds everyone that November's race has shifted from "whose parameters are bigger" to "how much work each token's money buys."

(Data from official announcements and public reports, 2026-11-05; vendor claims and third-party results are marked separately, with unverified items noted in the text. Model capability affects selection, but actual pricing follows the model plaza in real time.)

相关文章

GPT-6.2 Cuts Prices 20%: What the AI Price War Means

GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer

What Is Prompt Caching? How Cache Tokens Cut API Bills

Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi

All 12 Claude Models Compared: Price and Best Use

Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload

Which AI Model Writes Code Best? Speed and Price Tested

gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi

One API Key, Every Model: How to Integrate a Multi-Model Marketplace

Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa