Quanta AI Labs

07 / Feb 9 – Feb 15, 2026 · 36 RELEASES

AI News Week 7, 2026

Gemini 3 Deep Think and GLM-5 (745B) push reasoning frontiers. GPT-5.3-Codex-Spark delivers 1000+ tok/s real-time coding. Immersive media explodes with Alive, Hibiki-Zero, MOSS-TTS, and Qwen-Image-2.0.

Language Models (6)

  • Gemini 3 Deep Think

    Google DeepMind · Reasoning

    Major upgrade to Google's specialized reasoning mode, advancing complex problem-solving in science, research, and engineering domains.

  • GLM-5

    Zhipu AI · MoE Architecture

    745B MoE model (44B active) with native Agent Mode, MIT-licensed. Achieves record-low hallucination rate using the SLIME RL technique, rivaling top Western models.

  • GPT-5.3-Codex-Spark

    OpenAI · Code Generation

    First real-time coding model delivering 1000+ tokens/sec on Cerebras hardware. Smaller sibling of GPT-5.3-Codex optimized for ultra-low latency interactive coding.

  • MiniMax M2.5

    MiniMax · Code Generation

    Production LLM trained with extensive RL across complex real-world environments. Achieves 80.2% on SWE-Bench Verified, 100 tok/s at ~$1/hour continuous usage.

  • LLaDA 2.1-mini

    InclusionAI · MoE Architecture

    16B MoE diffusion language model with error-correcting token editing. Features Speed and Quality modes for configurable inference trade-offs.

  • Schematron-3B

    Inference.net · Small Language Models

    Fine-tuned 3B model for converting noisy HTML into schema-conformant JSON. Achieves 100% valid JSON output and improves factuality by ~10x when paired with web search.

Vision & Image (3)

  • DeepSeek-OCR-2

    DeepSeek · OCR & Document

    Advanced OCR system using visual causal flow methodology for human-like text recognition. Converts documents to markdown with layout preservation and streaming output.

  • DeepGen 1.0

    DeepGen Team · Text-to-Image

    Lightweight 5B multimodal model (3B VLM + 2B DiT) unifying image generation, editing, reasoning generation, and text rendering. Competitive with models 3–16x larger.

  • Qwen-Image-2.0

    Alibaba Qwen · Text-to-Image

    Next-gen 7B image model unifying text-to-image generation and editing. Native 2K resolution with professional typography rendering supporting 1K-token prompts.

Video & Animation (5)

  • DuoGen

    NVIDIA · Text-to-Video

    Framework for interleaved multimodal generation fusing pretrained MLLM with video DiT. Enables step-by-step guides, visual planning, and visual drafts for reasoning.

  • Alive

    ByteDance · Text-to-Video

    Unified audio-video generation system creating synchronized content from text or images. Features temporally-aligned cross-modal fusion for precise lip-sync.

  • JUST-DUB-IT

    Research · Video Editing

    Video dubbing via joint audio-visual diffusion. Adapts a foundational audio-video model for video-to-video dubbing using a lightweight LoRA approach.

  • Decart Avatar Live

    Decart AI · Real-time Video

    Realtime API that animates portrait images with audio input via WebRTC. Generates talking avatars at 25 FPS, 1280x720 with synchronized lip movements.

  • Delulu Stream

    Delulu · Real-time Video

    AI-powered streaming tool letting chat members modify broadcaster appearance in real-time — outfit swaps, character transforms, and environment effects linked to tips.

Audio & Speech (7)

  • SoulX-Singer

    Soul AI Lab · Voice Cloning

    Zero-shot singing voice synthesis model supporting melody-conditioned (F0) and score-conditioned (MIDI) control for high-fidelity voice cloning of unseen singers.

  • SoulX-Singer Demo

    Soul AI Lab · Voice Cloning

    Interactive demo page showcasing zero-shot singing voice synthesis with melody and score control, featuring audio samples across multiple vocal styles.

  • Hibiki-Zero

    Kyutai · Speech-to-Text

    Real-time 3B multilingual speech-to-speech translation model. Converts French, Spanish, Portuguese, and German speech to English with voice preservation. MIT licensed.

  • Hibiki-Zero Blog

    Kyutai · Speech-to-Text

    Technical deep-dive into Hibiki-Zero's real-time speech translation architecture, covering voice preservation, minimal latency design, and multilingual capabilities.

  • MOSS-TTS

    OpenMOSS · Text-to-Speech

    Open-source speech generation suite with 5 models: TTS, dialogue, voice generator, realtime, and sound effects. Supports 20 languages with high expressiveness.

  • MioTTS-2.6B

    Aratako · Text-to-Speech

    Lightweight LLM-based TTS model for English and Japanese with zero-shot voice cloning. Achieves 0.135 RTF on consumer GPUs at 25Hz framerate.

  • UniAudio 2.0

    Research · Text-to-Speech

    Unified audio language model with ReasoningCodec for both understanding and generation. Handles TTS, text-to-music, song generation, voice conversion, and speech recognition.

Agents & Automation (6)

  • VisionClaw

    Community · Agent Framework

    Real-time AI assistant for Meta Ray-Ban smart glasses combining voice interaction, visual perception via Gemini Live API, and automated actions via OpenClaw.

  • MiniMax Agent

    MiniMax · Agent Framework

    Desktop AI agent platform with Expert Agent feature for Mac, Windows, iOS, and Android. Automates coding, analysis, and productivity workflows at $19/month.

  • PicoClaw

    Sipeed · Agent Framework

    Ultra-lightweight AI assistant in Go running on <10MB RAM. Boots in <1 second on 0.6GHz processors, supports Telegram/Discord/QQ with web search and long-term memory.

  • Nanobot

    Nanobot AI · Agent Framework

    Open-source MCP host for building AI agents by combining MCP servers with LLMs. Configurable via YAML with web UI, targeting Slack/SMS/email integrations.

  • OnSpace AI

    OnSpace · Agent Framework

    No-code agentic app builder generating cross-platform mobile and web apps from text prompts or Figma designs. Integrates Sora 2, Gemini, and Supabase.

  • Dash

    Agno AGI · Agent Framework

    Self-learning data agent converting natural language to SQL with 6 layers of context grounding. Learns continuously without retraining from error patterns and fixes.

Developer Tools (7)

  • The Companion

    The Vibe Company · AI Infrastructure

    Web UI for managing parallel Claude Code and Codex sessions. Features streaming output monitoring, permission gating, and session persistence via WebSocket.

  • Daggr

    Hugging Face · AI Infrastructure

    Open-source Python library for building AI workflows connecting Gradio apps, ML models, and custom functions with auto-generated visual canvas for inspection.

  • verl

    ByteDance · AI Infrastructure

    Production-ready RL training library for LLMs supporting PPO, GRPO, RLHF with FSDP/Megatron backends. Scales to 671B parameters across hundreds of GPUs.

  • SLIME

    THUDM · AI Infrastructure

    LLM post-training framework integrating Megatron with SGLang for RL scaling. Powers GLM-5, physics reasoning (P1), and kernel generation (TritonForge).

  • Monty

    Pydantic · AI Infrastructure

    Minimal secure Python interpreter written in Rust for AI-generated code. Microsecond startup, snapshotting, and sandboxed execution for PydanticAI code-mode.

  • ONNX Model Zoo

    ONNX · AI Infrastructure

    Curated collection of pre-trained models in ONNX format covering vision, NLP, and generative AI. Historical archive now bridging to Hugging Face ecosystem.

  • Paper to Notebook

    Community · AI Infrastructure

    Web tool converting research paper PDFs into executable PyTorch notebooks using Gemini. Analyzes structure, designs architecture, generates code, and validates.

Research & Safety (2)

  • Awesome AI Agent Papers

    VoltAgent · Survey & Overview

    Weekly-updated curated collection of 224+ arXiv papers on AI agents covering multi-agent systems, memory/RAG, evaluation, tooling, and security.

  • PagedAttention (vLLM)

    UC Berkeley · Research Paper

    Foundational paper on efficient memory management for LLM serving using PagedAttention, enabling near-zero waste in KV cache memory and powering the vLLM engine.

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.