OpenAI GPT-Live to Alibaba Qwen3.8-Max: Wave of Next-Generation AI Model Releases
모델 출시 | Wed Aug 05 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 6 sources
New AI models spanning real-time voice AI, ultra-large open-weight models, and on-device agents were released.
Analysis
[OpenAI] unveiled real-time voice AI system GPT-Live [1]
- Full-duplex voice model that can listen and speak simultaneously
- Removed turn detector for natural conversational flow
- Can delegate to higher-tier models like GPT-5.5 when needed
- Powers ChatGPT desktop app's computer control and agent orchestration features
[Alibaba] released ultra-large open-weight model Qwen3.8-Max [4]
- 2.4-trillion-parameter flagship
- its largest and most capable model
- Benchmark performance rivals or exceeds Anthropic Fable 5
- Ranks after Fable 5 and Opus series on the Arena leaderboard
- Weights to be released next week
[OpenAI] solved 10 open math and theoretical CS problems with next-gen model Astra [2]
- Spans high-dimensional sphere packing
- coding theory
- and lattice cryptography
- Total token cost around $2
- 000 at Sol API rates
- Each proof formalized as a Lean certificate
- Follow-up to the Erdős unit-distance conjecture counterexample
[OpenAI] cut GPT-5.6 series prices significantly [3]
- GPT-5.6 Luna price reduced by 80%
- GPT-5.6 Terra price reduced by 20%
- GPT-5.6 Sol Fast mode offers up to 2.5x speed at 2x price
- Sol contributed to 20% serving cost reduction and 15%+ token generation efficiency gain
[Liquid AI] released on-device agent model LFM2.5-2.6B [2]
- 2.6-billion-parameter model running locally on laptops and smartphones
- Supports tool use and multi-step workflows
- 220 tok/s on Apple M5 Max with under 2.5GB memory usage
- Agent performance competitive with 4x larger models
[DeepSeek] published recipe to run V4 Flash model on a single AMD MI300X [5]
- Serves 304B parameter checkpoint on a single GPU
- 168.6 tok/s single-stream decode
- 830 tok/s at 64-stream burst
- 256K context verified
- architecture supports up to 1M
- Modifications for MI300X's FP8 fnuz format and MoE routing
Sources
- [1] How we built a realtime system for responsive voice AI in six months - OpenAI Blog
- [2] Ten advances in mathematics and theoretical computer science - OpenAI Blog
- [3] Building abundant intelligence - OpenAI Blog
- [4] China’s Alibaba takes another swipe at America’s AI supremacy - The Verge AI
- [5] DeepSeek V4 Flash on a Single AMD MI300X - Hacker News
- [6] Deploy local agents everywhere with LFM2.5-2.6B - Hugging Face Blog