Neurodigest Weekly (#125)
Neurodigest Weekly (#125)
A roundup of the latest news in LLM and generative models, including new releases and price changes.
Neurodigest Weekly (#125)
LLM
- Kimi K3 Weights — Moonshot has released the weights of the largest open model with 2.8T parameters (104B active). A system with 2TB of memory is required for local deployment.
- Price Reduction on GPT 5.6 — OpenAI has lowered prices: Luna is now five times cheaper ($0.2/$1.2), and Terra has dropped by 20%. The flagship Sol has launched a Fast mode (2.5 times faster at 2x the price).
- DeepSeek V4 Flash — The fresh version 0731 in the API surpasses GLM 5.2 and is compatible with Codex, remaining cheaper than Luna both in tokens and when calculated per task. The V4 Pro release is expected in early August.
- Astra from OpenAI — The new multi-agent model has solved 10 open mathematical problems using Lean, spending less than $2000 on tokens at Sol's rates.
Generative Models
- Seedance 2.5 — The video generator from ByteDance can now create videos up to 180 seconds long, has integrated with Blender, and offers point editing, but has become the most expensive on the market ($6 for 30 seconds).
Why it matters
AnalysisThis digest provides crucial insights into the latest advancements in LLM and generative models, which could significantly influence the future development of technologies in this field.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
AI Radar: GPT-6.1 Terra, Gemini 4 Carbon, Mistral Large 4 and OpenAI/Anthropic pricing changes
A roundup of recent AI news: new LLM releases, OpenAI and Anthropic subscription changes, world models, video generation, and research updates.
MIT researchers show GPT-6 Astra can improve other AI models by editing their weights
In an experiment, Astra raised Qwen3.5-0.8B's accuracy on MATH-500 from 7.9% to 48.1% by changing only 0.016% of the model's weights.
Part I. AI Agent Architecture: The Foundations
The eleven foundational building blocks of modern autonomous agents: model taxonomy, the agentic runtime loop, advanced prompting, distillation, skills architecture, memory systems, function calling, and FinOps.
Part II. AI Agent Architecture: From Model to Memory and Tools
A deep engineering dive: tokenization, KV-cache, sampling determinism, the training stack, quantization, context engineering, the safety stack, production RAG, ReAct, and multi-agent swarm systems.