Llama-modchip Enables 284B DeepSeek on 64GB RAM Without GPU
A Hacker News post highlights Llama-modchip, claiming it runs a 284B parameter DeepSeek model on 64GB of RAM without a GPU. If verified, this technique would significantly lower the barrier for local deployment of large language models.
Summarized by AI from one outlet
Source
- Hacker NewsRead original ↗
Latest
- Models
Meta Releases Lightweight Agent Model Muse Ember
Meta released Muse Ember, a lightweight model optimized for agent execution. With input costs of 10 cents per million tokens, it is significantly cheaper than frontier competitors like Haiku, aiming to offer a more cost-effective solution.
- Models
ByteDance releases Seedream 5.5 image model
ByteDance launched the Seedream 5.5 image creation model, now integrated into Doubao, Doubao Work, and Jimeng. The update significantly boosts generation speed, improves multi-round editing stability, and adds native support for transparent asset generation. The cost per 2K image drops to approximately 0.28 RMB, less than half of the previous Seedream 5.0 Pro.
- Models
Baidu Discloses RSI System Used for ERNIE 5.5 Flash Development
Baidu revealed its automated research system, ERNIE Evolve, which is used for developing the ERNIE 5.5 Flash model. The system autonomously proposes research directions and validates results, achieving state-of-the-art performance on benchmarks.
All news → · All rights belong to the original creators