AI Daily — 2026-11-05

English 中文

Open 300B agentic RL recipe, real-physics agent benchmark, and Qwen’s faster speculative decoding advance AI training, evaluation, and inference.

Covering 26 AI news items

🔥 Top Stories

1. FireFlash: Open Recipe for 300B Agentic RL at Pareto Frontier

FireFlash is a 300B-scale agentic RL recipe that reaches 46.2% on Terminal-Bench 4.0 while costing more than 30× less per task than frontier agents. The recipe trains on 64 NVIDIA B300 GPUs in under five days and combines curated RL environments, robust rollout recipes, and an optimized coding harness, with gains transferring to SWE benchmarks. Its main impact is reproducibility: teams get a documented path to frontier-adjacent long-horizon agents without frontier-scale spend. Source-reddit

2. FrontierPhysics Benchmark Tests AI Agents on Real Physics Research

FrontierPhysics introduces 56 end-to-end tasks authored by 42 physicists across nine domains, covering theoretical, computational, and experimental research. Unlike fixed-question or paper-reproduction benchmarks, it scores both research quality and correctness as agents execute full research workflows. The benchmark pushes agent evaluation toward genuine scientific utility and exposes where current systems still need human-grade judgment. Source-huggingface

3. Qwen Introduces SpecOPD for Faster Speculative Decoding

Qwen researchers propose SpecOPD, a branch-parallel on-policy distillation framework for speculative decoding that samples draft tokens and target supervision across independent branches. By packing those branches into proposal trees to share computation, the method targets longer acceptance lengths and faster LLM inference. If it holds up in production, SpecOPD could improve inference economics without changing model architecture. Source-reddit

📰 Featured

Open-Source Models & Developer Tools

CCORE Releases GLB-3D-Coder 4B Model and Full Weights — A full fine-tune of Qwen3.5-4B generates complete Three.js HTML scenes and exports GLBs for Godot, Blender, and Unity, with GGUF and Safetensors weights available. Its focus on reusable 3D assets makes it a practical open model for game and design pipelines. Source-reddit
Garry Tan Releases Open-Source gstack for Claude Code Agent Teams — The MIT-licensed toolkit provides 23 Markdown slash commands that simulate roles from CEO to QA and release engineer for coding agents. It reflects a growing pattern of packaging agent workflows as lightweight, forkable operating procedures. Source-github

Agentic RL & Embodied AI

TurnWise Introduces Turn-Level GRPO for Streaming Video RL — TurnWise replaces trajectory-level advantages with per-turn credit for streaming video understanding, aiming to avoid reinforcing redundant replies or penalizing useful follow-ups after failure. The approach addresses a key credit-assignment gap in multi-turn video agents. Source-huggingface
NEO-VLA Proposes Native Vision-Language-Action Policies Beyond Fixed Visual Encoders — NEO-VLA lets action policies access evolving visual computation rather than relying on a fixed encoder endpoint, arguing manipulation needs local geometry, boundaries, depth, and subtle state changes. The design points toward tighter vision-action coupling in robotics. Source-huggingface

World Models & Generative Video

GameVerse Dataset Enables Minute-Scale Long-Horizon Interactive World Modeling — GameVerse supplies minute-scale gameplay data and finer captions to train interactive world models beyond roughly one-minute coherence. Better long-horizon data could reduce drift in action-conditioned environment generation. Source-huggingface
MoneyPrinterTurbo Generates HD Short Videos From a Topic — This open-source tool automates scripts, media matching, subtitles, music, and final composition from a keyword, with a WebUI and API. It lowers the barrier to automated short-video production while raising familiar concerns about synthetic media volume. Source-github

Extreme-Low-Resource Inference

Kimi K3 2.78T Model Runs on One CPU in 8GB RAM — A portable C99 implementation runs a 2.78-trillion-parameter model on one CPU with disk streaming and 8.24 GB peak RAM, producing byte-identical output at very slow speeds. The project is less a practical serving path than a striking demonstration of memory hierarchy engineering. Source-github

⚡ Quick Bites

MoE Training Handbook Released with Megatron-LM Runbook — A new handbook pairs MoE training guidance with a Megatron-LM runbook, giving practitioners a structured reference for large-scale mixture-of-experts training. Source-reddit
DeepSeek-V4.1-Flash Hits 200+ tok/s on Eight 2018 V100s — A practitioner reports DeepSeek-V4.1-Flash exceeding 200 tokens per second on eight 2018-era V100 GPUs, highlighting how older accelerators remain viable for optimized inference. Source-reddit
Z.ai Reveals GLM-5.5 Training Pipeline and 32B/122B/256B Models — Z.ai’s chief scientist detailed GLM-5.5’s training pipeline and model sizes spanning 32B, 122B, and 256B parameters. Source-reddit
Delta Inference Engine v0.93 Boosts DeepSeek-V4.1-Flash on RTX 3090 — Delta Inference Engine v0.93 improves DeepSeek-V4.1-Flash performance on a single RTX 3090, according to a community release. Source-reddit
Factorized MLLM Embeddings Improve Composed Image Retrieval — A new paper uses factorized multimodal LLM embeddings to improve composed image retrieval, where queries combine reference images and text edits. Source-huggingface
PR-Agent: Open-Source AI Code Review Agent, Community-Maintained Legacy Project — PR-Agent is positioned as a community-maintained, open-source AI code-review agent for automated pull-request feedback. Source-github
Horizon Launches AI-Powered News Radar for Daily Bilingual Briefings — Horizon is an AI-powered news radar that produces daily bilingual briefings. Source-github
Scenario Labs Releases Agent Skills for AI Content Creation and DCC Tools — Scenario Labs released agent skills for AI content creation and digital content creation tools, packaging reusable capabilities for creative workflows. Source-github
RustingStrata vs llama.cpp benchmarks across Qwen GGUF models — A community benchmark compares RustingStrata and llama.cpp across Qwen GGUF models, adding evidence on runtime performance trade-offs. Source-reddit
Kokon adds editable layers around Qwen-Image 2.1 editor — Kokon adds editable layers around the Qwen-Image 2.1 editor, improving layer-level control in image-generation workflows. Source-reddit
Qwen3.8 Flash-Next Hits 250 tok/s on Single RTX 5080 With Basalt — A guide reports Qwen3.8 Flash-Next reaching 250 tokens per second on a single RTX 5080 with Basalt. Source-reddit
Qwen LLM’s Funny Internal Monologue on Tool Use and Vision — A community post highlights a Qwen model’s amusing internal monologue around tool use and vision, underscoring how reasoning traces can surface odd but revealing behavior. Source-reddit
Reddit User Praises OrcaCyber-Nano-1.0 Fine-Tune for Coding and Planning — A Reddit user praises the OrcaCyber-Nano-1.0 fine-tune for coding and planning tasks, adding anecdotal evidence for small specialized models. Source-reddit
DeepSeek Fixes HDMI Display After Qwen Fails in Local LLM Test — A local LLM test anecdote claims DeepSeek fixed an HDMI display issue after Qwen failed, an example of models being used for practical troubleshooting. Source-reddit
NVIDIA RTX PRO 5500 Blackwell Listed at $17,999, Above RTX PRO 6000 — NVIDIA’s RTX PRO 5500 Blackwell has appeared at $17,999, pricing above the RTX PRO 6000 and raising questions about workstation GPU positioning. Source-reddit
Reddit User Claims Qwen3.8-27B Quality Degraded After Initial Release — A Reddit user claims Qwen3.8-27B quality degraded after its initial release, highlighting persistent community scrutiny of hosted model updates. Source-reddit

Generated by AI News Agent | 2026-11-05