AI Daily — 2026-11-03

English 中文

Colibri enables local frontier MoE models in pure C, Karpathy's $48 nanochat trains GPT-2-level LLMs, and Qwen commits future versions to open source.

Covering 23 AI news items

🔥 Top Stories

1. Karpathy Releases nanochat: Train GPT-2-Level LLM for $48

Andrej Karpathy’s nanochat packages tokenization, pretraining, finetuning, evaluation, and inference into a minimal single-GPU-node harness. With a single --depth parameter, it auto-configures compute-optimal models and can train a GPT-2-capability model for roughly $48, or about $15 on spot capacity. The release lowers the practical barrier to end-to-end LLM experimentation and gives practitioners a compact reference for cost-aware training. Source-github

2. Qwen Confirms Future Qwen 4.0 and 5.0 Will Remain Open Source

Qwen Head Liu Dayi Heng denied rumors that Qwen4 would be the final open-weight release and confirmed Qwen 4.0 and 5.0 will remain open source. The company frames AI as a foundational capability and cites 3B+ cumulative downloads, 460+ open-source models, and 300K+ derivative models. The commitment is significant for the open-weight ecosystem amid uncertainty about frontier labs’ release strategies. Source-reddit

3. Colibri Runs Frontier MoE Models Locally in Pure C

Colibri is a pure-C inference engine that runs very large mixture-of-experts models on consumer hardware by keeping active experts in RAM and streaming the rest from disk. It requires no GPU and supports GLM, Kimi, DeepSeek, Qwen, OLMoE, and other model families. If its performance holds, it could make frontier-scale MoE experimentation viable on ordinary machines and strengthen local-first LLM tooling. Source-github

📰 Featured

Video Generation & World Models

Wave Forcing Enables Speed-of-Light Streaming Video Generation — Wave Forcing co-designs generation, training, and execution for real-time long-video output, targeting serialized denoising and camera drift in chunk-wise autoregressive models. It points toward lower-latency, more stable streaming video generation. Source-huggingface
Agent Harness Bridges Video Generation and World Modeling — A new executable agent harness manages context, tools, state, feedback, and budgets to support controllable multi-shot video generation and interactive agent environments. The work connects agentic video generation with video world models. Source-huggingface

Agents & Computer Use

CUA-Master Introduces Recursive Improvement Framework for Computer-Use Agents — CUA-Master uses a nested double-loop recursive improvement framework to address limited diversity, controllability, and reliability in computer-use interaction data. It aims to move open-world agents beyond diminishing returns from self-generated experience. Source-huggingface
CoworkOS Introduces General-Purpose Harness for Long-Horizon Agent Coworkers — CoworkOS separates task strategy from execution management to support long-horizon productivity tasks requiring artifact delivery plus continuous monitoring. It targets more reliable real-world agent coworkers. Source-huggingface
Vercel Labs Releases Agent-Desktop CLI for AI Computer Use — agent-desktop lets agents read and operate desktop apps via OS accessibility trees on macOS, Windows, and Linux using JSON snapshots and actionable refs. It avoids pixel matching and supports headless operation, though it carries privacy warnings for sensitive desktop content. Source-github

Tools & Inference

Heretic Tool Modifies LLMs Without Training to Remove Censorship — Heretic uses contrasting prompts and scorers to steer locally runnable LLMs without training, suppressing refusals or altering tone and emotional alignment. It highlights the growing ease of behavior modification and the corresponding safety and governance questions. Source-github
Basalt 0.2.0 Boosts Qwen 3.8 Flash Next Inference on Blackwell GPUs — Basalt 0.2.0 adds NVFP4 and other quantization options for Qwen 3.8 Flash Next on Blackwell GPUs, reporting 300 t/s prose, 600 t/s structured output, and 10K t/s prefill on RTX 5090 plus 5060 Ti with 128GB RAM. It emphasizes quality comparisons against a BF16 llama.cpp reference alongside speed. Source-reddit

⚡ Quick Bites

Qwen Code Repo Leaks Qwen4-Max, Coder, and Omni Model IDs — A Qwen code repository appears to reveal model identifiers for Qwen4-Max, Coder, and Omni, fueling speculation about the next release lineup. Source-reddit
Jacobian Lens Reveals Qwen Reasoning Loops; Weight Edit Fixes Them — A Jacobian-lens analysis identifies reasoning loops in Qwen and shows that targeted weight edits can correct the stuck behavior. Source-reddit
RTX 5090 64GB and 96GB VRAM Mods Reportedly Confirmed by Repair Tech — A repair technician reportedly confirms 64GB and 96GB VRAM modifications for the RTX 5090, a potentially notable development for local LLM capacity. Source-reddit
llama.cpp Gemma Graph Optimizations Boost E2B/E4B Prompt Processing — New Gemma graph optimizations in llama.cpp deliver major prompt-processing gains for E2B/E4B models. Source-reddit
LLM Agents Enable New Paradigm in Computer Graphics Code Generation — A new paper argues LLM agents enable a distinct paradigm for generating computer graphics code. Source-huggingface
Paperclip Launches Open-Source Orchestration for AI Agent Teams — Paperclip releases open-source orchestration for coordinating AI agent teams. Source-github
Reflection product team hosts AMA on open-weight Beam 501B model — Reflection’s product team is hosting an AMA on its open-weight Beam 501B model. Source-reddit
Claude Rewrites Gemma 4 Neurons to Fix Italian Tense Translation — Claude was used to rewrite Gemma 4 neurons and fix an Italian tense-translation issue. Source-reddit
DeepSeek V4.1 Flash 1017 Interprets Small Neural Network Weights — DeepSeek V4.1 Flash 1017 can also perform weight-space interpretation for small neural networks. Source-reddit
Revolut Cuts AI Coding Costs 10x by Self-Hosting Instead of Claude — Revolut reportedly cut AI coding costs by 10x by self-hosting instead of using Claude. Source-reddit
Library of Layers Adaptively Orders Transformer Layers at Runtime — A Library of Layers approach adaptively orders transformer layers at runtime. Source-reddit
Users Ask What Local LLMs Run on 32GB RAM, 16GB VRAM — Local LLM users are discussing which models run on a normal 32GB RAM and 16GB VRAM setup. Source-reddit
Prototype AMD Radeon RX 470 Board With Four GPUs Pictured — A prototype AMD Radeon RX 470 board with four GPUs has been pictured. Source-reddit

Generated by AI News Agent | 2026-11-03