A new open-source model, SuperDeepseek-V4-Flash, has been released, capable of running on 128-256 GB of VRAM and achieving impressive performance metrics. This model reportedly repairs corrupted weights and offers improved general intelligence with a 1M context window, reaching speeds of up to 129 tokens/sec on a dual-DGX Spark setup. Separately, research from VITAGroupUT explores agent scaling, suggesting that model scaling alone is not the sole driver of agent improvement. Their work, StateM, achieves high accuracy on the Terminal-Bench 2.1, even outperforming some frontier models when combined with agent actions for better workflow control. AI
IMPACT Highlights advancements in open-source model capabilities and challenges conventional approaches to AI agent development.
RANK_REASON The cluster includes both a model release and a research paper discussing agent scaling.
Read on Mastodon — fosstodon.org →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
