DeepSeek, Qwen, Grok, and Liquid AI Unveil Wave of Next-Gen Frontier Models
모델 출시 | Thu Aug 13 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 4 sources
New model releases ranged from massive MoE architectures to edge-deployable vision-language models.
Analysis
[DeepSeek] released DeepSeek V4 Pro 0813 [1]
- New model accessible via OpenRouter
- Pro version of the DeepSeek V4 series
[Qwen] deployed Qwen3.8-2.4T-A95B open model [2]
- MoE architecture with 2.4T total parameters and 95B activated
- Activates 10 routed experts plus 1 shared expert out of 512
- Context extends from 262
- 144 tokens by default up to approximately 1
- 010
- 000 tokens
- Improved performance across coding
- research
- and long-horizon agent tasks
- reasoning_effort controls reasoning depth while preserve_thinking maintains reasoning context
[xAI] released Grok 4.6 with enhanced agentic and coding performance [3]
- Tied with GPT-5.6 Sol at 61 points on the Artificial Analysis Intelligence Index
- Enhanced long-horizon agentic and visual/interactive tasks compared to Grok 4.5
- Immediately available in Cursor and Grok Build with 2x usage during the first week
- Trained with domain-specialized agentic RL for kernel optimization
- web development
- CAD
- and more
- Priced at $2 input / $6 output per 1M tokens
[Liquid AI] launched LFM2.5-VL-3B vision-language model for edge devices [4]
- 3.1B parameter VL model capable of on-device execution
- Enhanced screen/UI understanding
- grounding
- multi-image input
- and function calling
- Combines SigLIP2 400M NaFlex vision encoder with LFM2.5-2.6B backbone
- Pretrained on approximately 34T tokens with 4x expanded vision data
- Vocabulary doubled to 128K to support non-Latin scripts