AI Daily — 2026-11-02

English 中文

Kronos launched an open-source financial K-line foundation model, Qwen pledged open-source 4.0/5.0, and ISTA released a 763B one-bit DeepSeek coder.

Covering 23 AI news items

🔥 Top Stories

1. Qwen Confirms Qwen 4.0 and 5.0 Will Remain Open Source

Qwen Head Liu Dayi Heng officially denied rumors that Qwen4 would be the last open-weight release, confirming that Qwen 4.0 and Qwen 5.0 will continue to be open source. The commitment matters because Qwen reports over 3 billion cumulative downloads, 460+ open-source models, and 300,000+ derivative models, making its roadmap a bellwether for the open-weight ecosystem. It also reinforces competitive pressure on closed labs to justify proprietary access. Source-reddit

2. ISTA Releases 763B DeepSeek-V4.1-Flash Coder at About 1 Bit per Parameter

ISTA released a Coder build of DeepSeek-V4.1-Flash-1017 compressed with GSQ and RCO, pruning half of its routed experts and keeping 192 of 384 per layer. The result is 72.1 GB of transformer weights at roughly 1.05 bits per original transformer parameter, with the shared Engram file and vision projector unchanged. This is a notable stress test for extreme MoE compression and local deployment of very large coding models. Source-reddit

3. Open-Source Sparse Attention Enables 10M-Token Context Cheaply

A fully open-source sparse attention method sparsifies Qwen 3.8 27B attention with no observed accuracy loss and can be trained natively on next-token prediction to match full-attention validation loss. The team reports context extension to 10M tokens with degradation only on specific tasks while NIAH benchmarks remain fine, with public kernels and SGLang integration. If reproducible, this lowers the cost of long-context inference on local hardware for small open-weight models. Source-reddit

📰 Featured

Open Source Models & Infrastructure

Kronos: Open-Source Foundation Model for Financial K-Line Data — Kronos releases an open-source decoder-only financial K-line foundation model with a hierarchical OHLCV tokenizer and AAAI 2026 acceptance, giving quant researchers a reproducible pretrained baseline. Source-github
FreeToken Runs 290B+ MoE Models Locally on Consumer Hardware — FreeToken is an edge-native MoE serving engine that unifies GPUs, CPUs, host memory, and interconnects with bandwidth-adaptive co-execution and expert caching for frontier-scale local inference. Source-github

Multimodal & Representation Learning

FrameForge Uses Video Foundation Model Priors for Unified Image Generation — FrameForge introduces unified conditional image generation that leverages pretrained video foundation models and handles varying condition-target relationships across tasks. Source-huggingface
Theorem Proves JEPAs Learn Shared Platonic Representations Across Modalities — Researchers prove that separately trained Joint-Embedding Predictive Architectures learn the same representation up to rotation across modalities, offering a theoretical guarantee for multimodal alignment. Source-huggingface

Agents, Embodied AI & Security

R^2Orch: Recursive Self-Play for Agent Orchestration — R^2Orch proposes a recursive self-play orchestration paradigm with an inner self-play loop to improve agent skill learning and address hand-designed curricula, advantage collapse, and noisy per-skill evaluation. Source-huggingface
REA: Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control — REA pairs slow foundation-model reasoning with fast reactive policies for asynchronous embodied control, targeting incompatible evaluation stacks and deliberative-reflexive coordination. Source-huggingface
Open-Source AI Security Operations Center with Auditable LLM Agents — AiSOC is a self-hostable AI security operations center with auditable LLM agents, alert fusion, MITRE ATT&CK investigation, a replayable decision ledger, and an MCP server for Claude, Cursor, and Continue. Source-github

⚡ Quick Bites

Anthropic Tightens Claude Code Limits to Block Reverse Engineering and Game Cloning — Anthropic reportedly narrowed Claude Code usage limits to prevent reverse engineering and game cloning, renewing self-hosting debates. Source-reddit
LLMs Can Natively Read and Write Neural Network Weights — A new weight-space programming demo suggests LLMs can inspect and edit neural network weights directly. Source-reddit
Co-Designing System-1 and System-2 for Scalable AGI — A new paper proposes co-designing fast and slow reasoning components for scalable AGI architectures. Source-huggingface
Handy: Open-Source Offline Speech-to-Text App for Desktop — Handy offers local, offline speech-to-text transcription for desktop users. Source-github
Kornia: Differentiable Computer Vision Library for PyTorch Spatial AI — Kornia provides differentiable computer vision operators for PyTorch spatial AI workflows. Source-github
Hex Offers Local AI Voice Dictation Across Desktop Platforms — Hex delivers local AI voice dictation across desktop platforms. Source-github
AI Engineering Hub Offers 93+ LLM, RAG, and Agent Projects — AI Engineering Hub collects more than 93 practical LLM, RAG, and agent projects for builders. Source-github
Qwen 3.8 27B Uses Blender Python to Speed Math Reasoning — A user reports that Qwen 3.8 27B sometimes uses Blender Python to aid math reasoning. Source-reddit
Local Quants vs Claude Tested on 9 Real CUDA Bugs — A benchmark compares local quantized models with Claude on nine real CUDA bug-fixing tasks. Source-reddit
Amazon Allegedly Cancels GPU Orders, Closes Accounts for Compute Hardware Purchases — Reports allege Amazon canceled GPU orders and closed accounts tied to compute hardware purchases. Source-reddit
GPT-6 Astra Edits Qwen Weights, Boosts MATH-500 to 48.1% — A report says GPT-6 Astra edited Qwen weights and raised MATH-500 performance from 7.9% to 48.1%. Source-reddit
Community Proposal for Private LLM Quality Benchmarks Over Tok/s — A community proposal argues private LLM evaluations should prioritize quality benchmarks over tokens-per-second metrics. Source-reddit
Reflection Lab to Host AMA on Open-Weight Beam Model — Reflection Lab will host an AMA on its open-weight Beam model. Source-reddit

Generated by AI News Agent | 2026-11-02