Quanta AI Labs

42 / Oct 5 – Oct 11, 2026 · 38 RELEASES

AI News Week 42, 2026

The Cheap Tier Got Good. Anthropic released Claude Haiku 5.5 on 7 October at a tenth of Haiku 4.5's API price, and its own table shows it jumping from 15.7% to 72.4% on OSWorld. Google shipped Nano Banana 2.1 the day before at about half the price of Nano Banana 2, along with EmbeddingGemma 2, which now embeds text, images, audio and video in one 740M model. OpenAI pushed GPT-6 to every ChatGPT tier on 7 and 8 October, with free users getting GPT-6 Luna and answers that can include interactive charts and forms. The research story was OpenAI publishing 719 mathematics manuscripts written by an unreleased model, about 42% of them formalized in Lean, then withdrawing three the next day over a sign error. Google DeepMind open-sourced AlphaProtein Novo for designing new enzymes, and Kandinsky Lab released Kandinsky 6.0 under MIT, generating video with synchronized audio at 3B and 29B. A quieter trend ran through the agent tools: Hex-Rays shipped an official IDA MCP server, Ghidra MCP reached a 7.0 release candidate and REA added Android support, so coding agents can now drive the main disassemblers directly, and the most-starred new projects were again agent skills for slides, design, marketing and video.

Language Models (5)

  • Claude Haiku 5.5

    Anthropic · Small Frontier Model

    Anthropic's fastest model, released 7 October after Claude Opus 5.5 (week 40) and Sonnet 5.5 (week 41), and the first Haiku with effort levels. API pricing is $0.10 input and $0.50 output per million tokens for prompts up to 100K tokens, a tenth of Haiku 4.5's $1 and $5, rising to $0.50 and $2.50 above that. Against Haiku 4.5, Anthropic reports 72.4% on OSWorld versus 15.7%, 39.2% on Terminal-Bench versus 0.0%, and 1,620 versus 735 on GDPval-AA, all vendor-reported. Anthropic still recommends Sonnet 5.5 and Opus 5.5 for complex agentic coding and positions Haiku for narrow tasks and subagent work. Available on the Claude Platform, AWS, Google Cloud and Azure as claude-haiku-5-5.

  • GPT-6 for Everyone in ChatGPT

    OpenAI · ChatGPT Rollout

    OpenAI began rolling GPT-6 out across ChatGPT worldwide on 7 October, reaching Free and Go users on 8 October. Plus, Pro, Business and Enterprise get GPT-6 Sol (covered here in week 40), while Free and Go get GPT-6 Luna. The rollout adds Intelligent UI, where answers can include interactive charts, forms, tappable buttons and diagrams on the web and updated mobile apps. OpenAI says GPT-6 Instant starts web-search answers 44% sooner on average than GPT-5.6 Instant. GPT-6 replaces Latest in the model picker, GPT-5.6 Sol stays selectable, and Work and Codex models are unchanged. openai.com blocked automated reading, so details come from Unite.AI and AndroidHeadlines coverage, which agree on the tier split and the 44% figure.

  • EmbeddingGemma 2

    Google DeepMind · Multimodal Embeddings

    Released 6 October as the follow-up to EmbeddingGemma 300M (covered here in week 32 of 2025), now embedding text, images, audio and video in one space. The text model is 270M parameters; adding the optional 170M vision encoder and 300M audio encoder makes a 740M multimodal model. An 8K-token window holds up to 5.5 minutes of audio, 29 images or 58 video frames. Vectors are 768 dimensions and can be cut to 512, 256 or 128 through Matryoshka Representation Learning. Google reports MTEB Code rising to 78.68 from 68.76, vendor-reported, and says it keeps the original's multilingual strength. Apache 2.0.

  • Reka Rho-1

    Reka AI · Omni Reasoning Model

    A 19B omni-reasoning model trained from scratch and released as a research preview on 5 October. One network understands and generates text, images and video, reasons over them and emits robot actions, all as tokens in a shared KV cache with no tool calls or second model. Each transformer block splits compute into an understanding stream and a diffusion generation stream. Reka reports base video generation at 0.79x real time, and a distilled Flash variant that cuts denoising from 99 steps to 8 and returns a 5.3-second clip in about one second, from its own testing. Video is capped at 672x384. No public weights or API; access to the preview is by request. Unrelated to Microsoft's 2024 Rho-1 language model.

  • ALoDLM-8B

    Amazon · Looped Diffusion LLM

    Amazon's Adaptively Looped Diffusion Language Model, in 1.7B and 8B sizes on Qwen3 backbones. Instead of giving every token the same compute, it runs shared transformer layers repeatedly: tokens it is confident about commit early, harder ones keep their latent state for more passes, and a learned halting policy decides when to stop. In the paper's 11-benchmark table the 8B averages 80.3 against Qwen3-8B's 78.5, with 87.8 on HumanEval and 49.5 on GPQA-Diamond, vendor-reported. Ships with an inference engine based on nano-vLLM. CC-BY-NC-4.0, non-commercial. Paper arXiv 2610.04198 (3 October), weights uploaded 2 October.

Vision & Image (3)

  • Nano Banana 2.1

    Google DeepMind · Image Generation Model

    Google's new image generation and editing model, announced 6 October and built on Gemini 3.6 Flash, following Nano Banana 2 Lite (covered here in week 27). Google says it improves visual design, mask-based editing and keeping subjects consistent across edits. The model card lists up to 1M input tokens with 4K tokens of image output and 64K of text output, and a thinking-mode text-to-image preference Elo of 1050 against 990 for Nano Banana 2 and 935 for Nano Banana Pro, vendor-reported. Coverage puts the price at about half of Nano Banana 2. It is rolling out to the Gemini app, AI Studio, the Gemini API, Search AI Mode, Flow, Stitch and Google Ads. Google lists small text and character consistency as known weaknesses.

  • Iris-3B

    Speridlabs · Pixel-Space Diffusion

    A 3B-parameter text-to-image model, released 7 October, that generates pixels directly with no VAE or latent space, at around one megapixel. The architecture uses 8 dual-stream and 16 single-stream blocks over 16x16 pixel patches. At 1024x1024 it scores 0.798 on GenEval, 86.52 on DPG, 0.857 on LongText and 0.540 on OneIG, vendor-reported, and the authors claim it is competitive with latent models at a fraction of the training compute. Fine-tuned without architecture changes, the same model also does depth estimation and image restoration and upscaling. Weights, code and a demo are public under Apache 2.0; paper arXiv 2610.09450.

  • ArtCraft Crafting Apps

    ArtCraft (Storyteller) · Open Adobe Alternatives

    The team behind ArtCraft, an open-source AI image and video studio, launched seven free native desktop apps written in Rust on 7 October, each a clean-room take on an Adobe tool: PhotoCraft (Photoshop, with PSD support), VectorCraft (Illustrator), FilmCraft (Premiere Pro), LightCraft (photo library and RAW), a PDF workbench, EffectCraft (motion graphics) and DesignCraft (page layout). They run fully locally, most are early alpha or still in development, and the repos are Apache-2.0 on GitHub, where PhotoCraft gathered 42,961 stars in 11 days. The apps are not AI features themselves; the AI link is ArtCraft's own studio and press reports that they were built with AI assistance.

Video & Animation (3)

  • Kandinsky 6.0 Video

    Kandinsky Lab · Open Audio-Video Model

    An open-source family of diffusion models, released 6 October, that generate 5-second clips with synchronized 44 kHz audio, including lip-sync, from text or an image plus text. It comes in Lite (3B) and Pro (29B). A dual-stream CrossDiT joins the Kandinsky 5.0 video stream to a newly trained audio stream through bidirectional cross-attention, and a plug-in super-resolution model lifts output to 1920x1080; Pro defaults to 864x480 at 24fps for 125 frames. MIT, with pretrained, full and distilled checkpoints, a Hugging Face Space, and day-one support in Diffusers, ComfyUI, vLLM-Omni, SGLang and FastVideo. Kandinsky Lab reports Pro beating 5.0 Pro in human evaluation, self-reported. Kandinsky Lab's homepage was covered in week 47 of 2025.

  • /brag

    Latent Spaces · Launch Video Skill

    Since Week 27 2026: the agent skill that turns a software project into a short launch video with music and share copy added voiceover support in July, moved to v0.3.0 in September and gained official Codex plugin packaging on 29 September. A new /brag-slim builds the video directly with Claude Opus 5.5 without HyperFrames and runs by default on that model, while --full keeps the HyperFrames path. Runs stay local and skip secrets, and letsbrag.app makes a 20-second video from a pasted URL. There are no tagged GitHub releases, so versions come from commit messages. MIT, 15,018 stars.

  • Animate

    cth9191 · Procedural Animation Skill

    A Claude Code plugin, created 4 October, that makes short animated videos entirely in code on one deterministic canvas, rendered to MP4 with headless Chromium and ffmpeg. Before animating it asks for approval at three checkpoints: a story beat table, style frames and a storyboard. It ships seven styles (cut paper, crosshatch ink, riso, sketchbook, a manim-like math style, pixel and isometric) and can match a new style from reference media. It beat-maps your own music, times ElevenLabs voice-overs with faster-whisper, screenshots real product sites and exports 9:16, 1:1, 4:5 and 16:9 from one piece. Needs Claude Code, Node 18, Playwright and ffmpeg. MIT, 436 stars.

Audio & Speech (2)

  • WorldSonus

    NoizAI · Real-Time Video-to-Audio

    An interactive video-to-audio model, posted 6 October with code and weights, that adds synchronized stereo sound to generated worlds and ordinary videos. A causal autoregressive diffusion model produces audio in 100 ms chunks from three video frames each, at a reported real-time factor of 0.41, and text instructions can change mid-sequence through chunk-indexed prompt scheduling. Stereo spatial alignment comes from training on curated stereo and ambisonic data, with 48 kHz stereo output in offline or streaming mode. Visual features come from DINOv3 and text from T5Gemma 2. The authors say it matches or beats bidirectional models on open-domain video-to-audio benchmarks, self-reported. Weights are CC BY-NC 4.0, non-commercial.

  • VoiceStudio v0.5.6

    debpalash · Local Voice Studio

    Since Week 37 2026: the local, open-source ElevenLabs alternative for voice cloning, voice design, dubbing, dictation, transcription and audiobooks in 646 languages, whose site is voicestudio.sh, moved to a redesigned Electron desktop app in v0.5.3. v0.5.5 lets a cloned voice read a script in another language, and v0.5.6 (23 September) answers Twilio phone calls in a saved voice, connects to Claude Code, Cursor, Codex CLI and the OpenAI Agents SDK over MCP, and finishes long audiobook chapters on 8 GB GPUs. AGPL-3.0, with optional paid Pro licences. 57,537 stars, up from 24,672.

3D & Spatial (4)

  • WorldPlay2

    WorldPlay2 team · Interactive World Model

    A real-time interactive world model built on Wan2.2 I2V-A14B, steered with frame-aligned actions (camera pitch and yaw, movement, jump) plus structured captions that separate scene, character identity and events such as lighting a stove. A learned history compressor turns past frames into memory tokens about 32x shorter than the full context for long-horizon consistency, and a distillation method called Stable Forcing yields a 4-step autoregressive Fast model alongside 40-step variants. The paper was posted 28 September; weights and inference code arrived 6 October and a browser demo on 7 October. Comparisons against WorldPlay and LingBot-World-V2 are self-reported. Checkpoints are CC BY-NC 4.0, non-commercial.

  • GenIA

    Meta Reality Labs and University of Tübingen · Input-Aligned 3D Reconstruction

    A method, posted 8 October (arXiv 2610.12388), that makes Meta's frozen SAM 3D model match real photos and video at generation time without retraining. It takes object translation and scale from observed geometry while keeping SAM 3D's learned rotation estimate, and aligns appearance with visibility-biased attention, fusion across views and differentiable-rendering guidance during denoising; an optional last step fine-tunes small adapters. It handles single images, multiple views and dynamic input, and on dynamic sequences costs about 13 times SAM 3D's single-image runtime against about 160 times for competing methods, author-reported. Andreas Geiger is among the authors. Builds on SAM 3D (week 47 of 2025).

  • DistScene

    Kunming Luo, Ping Tan et al. · 3D Scene Generation

    Generates complete 3D scenes from images, producing the environment and the objects in it together. Its training data is made automatically: TRELLIS.2 (covered here in week 51 of 2025) creates 3D objects and environments, physics checks guide how they are placed, and rendered views become the input images. The output scenes load directly into navigation simulators. Paper arXiv 2610.06960 (3 October) by 11 authors including Kunming Luo, Xuelong Li and Ping Tan. Preview assets are on Hugging Face; code, checkpoint and dataset are listed as coming soon.

  • CoDimRecon

    Carnegie Mellon University, Tsinghua SIGS and Genesis AI · Sim-Ready Scene Reconstruction

    An agentic system that turns multi-view photos into editable scenes ready for physics simulation, modelling soft objects at the right dimension: curves, surfaces or volumes. Telephone cords become elastic rods and plastic bags thin shells; the project shows paper holding a 140-degree fold and cushions that compress about 23 mm and recover. Paper arXiv 2609.36024 (28 September), with Minchen Li among the authors, from Carnegie Mellon, Tsinghua SIGS and Genesis AI. Code is listed as coming soon.

Agents & Automation (13)

  • FlashDexRetarget

    KAIST AI and Holiday Robotics · Dexterous Hand Retargeting

    Turns large collections of recorded human hand-object demonstrations into motions a dexterous robot hand can perform, training one reinforcement learning policy across all motions instead of one per demonstration. It reaches 90% retargeting success on a 50-motion XHand benchmark in 29 GPU-hours, which the authors put at about 100 times less compute than CHORD's 2,847, and converts more than 600 of 1,000 motions within 60 GPU-hours, author-reported. Tested on XHand and Sharpa Wave Hand on Isaac Sim 5.1 and Isaac Lab 2.3, with TACO, OakInk2 and HOT3D data pipelines. Paper arXiv 2610.01849 (October), MIT code.

  • TERRA

    EPFL (Mathis Group) · Musculoskeletal Locomotion

    Terrain-Aware Reconstruction, Retargeting and Control, from EPFL's Mathis group (the team behind MuscleMimic and Kinesis), teaches simulated muscle-driven human bodies to walk over terrain that is not flat. It infers the supporting surfaces from motion-capture data, applies biomechanical constraints when transferring the motion, and trains controllers on the result. The locomotion library totals 11.3 hours, 9.4 of them used for training, and the paper shows steep ramps, stairs, seats, separate platforms, narrow beams and uneven ground. Paper arXiv 2609.38653 (29 September); Apache-2.0 code on GitHub, updated through 9 October.

  • REA 6.4

    morluto · Reverse Engineering MCP

    Reverse Engineer Anything: one MCP server and CLI (npx rea-agents setup) that lets coding agents inspect software without its source code, covering native binaries through Hopper, Ghidra or IDA, JavaScript and Electron apps, websites, HAR captures, .NET assemblies, Android APKs, firmware and process behaviour, with evidence attached to each finding. Setup supports Claude Code, Codex, Cursor, Gemini CLI, Grok Build and others. Version 6.4.0 on 11 October adds Android device work over ADB, Apktool decoding, JEB project browsing and Ghidra support for Windows DLLs and MIPS ELF files, after 6.2.0 and 6.3.0 on 9 October. MIT, 77,249 stars and 16,510 forks, 35 releases since July.

  • Hex-Rays IDA MCP

    Hex-Rays · Reverse Engineering MCP

    The official, free and open-source MCP server from IDA's maker, letting AI agents disassemble and decompile with IDA 9.4 or later through idalib on the Pro, Home, Classroom and OEM editions. It uses a small code-mode toolset of database operations, IDAPython execution and API reference search, which Hex-Rays says uses about 20% fewer tokens than popular alternatives, self-reported. Built on IDA Nexus, it lets several agents and the IDA GUI work on one database at once. Installs through hcli or as plugins for Claude Code, Codex, Copilot CLI and Pi, over stdio or HTTP. Announced 25 September, with fix releases on 3 and 8 October. MIT, 949 stars.

  • Ghidra MCP 7.0 RC2

    bethington · Reverse Engineering MCP

    An MCP server that gives AI agents read and write control of Ghidra 12.1.4, as a GUI plugin or a headless or Docker server, exposing 209 tools for decompiling, cross-references, data-flow analysis, renaming, typing, creating structures, P-code emulation and live debugging. It enforces naming conventions, carries documentation across binary versions by function hash and connects to shared Ghidra servers. The 7.0.0-rc.2 pre-release on 9 October consolidates tools (53 removed and 9 added since rc.1), drops game-specific tools and turns debugger proxy tools off by default; the stable 7.0.0 is not out yet. Apache-2.0, 5,356 stars.

  • universal-modder

    rehan-remade · Game Modding Agent

    Skills, a um CLI and the fal MCP server that let AI coding agents mod PC games the user owns, installed as a plugin for Claude Code, Codex, Gemini CLI, VS Code Copilot, Cursor and OpenCode. Every job follows a ten-step loop: search the knowledge base, recon, set up a backed-up lab, read the real code, build one slice, generate art, 3D and audio with fal or a local ComfyUI, verify in the running game, record, package and write a field note that agents contribute back by pull request so the next agent starts from it. The rules ban redistributing game files, decompiled dumps and cheating. MIT, 6,425 stars, created 30 September and pushed through 10 October.

  • OpenRig v0.6.9

    mvschwarz · Persistent Agent Teams

    An open-source system, at openrig.dev, for building lasting teams of coding agents from Claude Code, Codex and Pi. A lead agent you talk to hands work to specialist seats with roles, shared context and owned work, tracks it and reports back; agents message each other with rig send, queue work with rig queue, and a terminal UI shows the team. It needs Node 22 or 24 and tmux. Three releases this week, ending with v0.6.9 on 11 October, fixed Slack integration, messaging across hosts and seat restoration. Apache-2.0, 6,763 stars, 30 releases.

  • Personal Jarvis 3.0

    Dodaios · Desktop Agent App

    An open-source local desktop app for Windows, macOS and Linux that brings several agent tools together. It runs Claude Code, Codex, OpenCode, Kimi Code, Grok Build, Cursor CLI and other coding agents side by side in terminal panes that you can steer by voice, and adds long-lived agents with their own memory and schedules, a wake-word voice assistant, desktop and browser control, a local wiki for memory and more than 40 plugins and MCP servers. It works with OpenAI, Anthropic, Gemini, OpenRouter or Ollama models, or existing subscriptions. v3.0.0 on 9 October labels plugins whose sign-in flow has not been fully tested as Preview. The repo moved from PersonalJarvis to Dodaios. Apache-2.0, 165 stars.

  • Papermorph

    DozenTwelve · PDF-to-Web-Book Skill

    An agent skill that turns a PDF book into an animated, narrated, interactive web book, running a pipeline from book plan to storyboards, narration, animation and quizzes. It installs with npx skills add and runs as /papermorph in Claude Code. The author built it with Opus 5.5 and reports that GPT-6.1 Sol also runs it well and Sonnet 5.5 works with rough edges. A live bookshelf holds two maths books and three physics chapters. Image models, multilingual output and background music are not supported yet. MIT, 668 stars, created 3 October and pushed through 11 October.

  • PPT Master v6.7

    hugohe3 · AI Native Decks

    An agent workflow that turns documents or topics into native, editable PowerPoint files, with real shapes, transitions and animations, data-backed charts and tables, narration from speaker notes and support for your own .pptx templates. v6.7.0, released 8 October, adds a browser review page for the design spec, makes Nano Banana 2.1 the default Google image model, accepts transparent backgrounds for GPT Image models, and fixes content that source conversion and PPTX or SVG round-trips used to drop silently. MIT, 59,655 stars.

  • Marketing Skills

    coreyhaines31 · Marketing Agent Skills

    Since Week 10 2026: Corey Haines' agent skill library for conversion optimisation, copywriting, SEO, analytics, paid ads, retention and go-to-market has grown to 50 skills, all building on a shared product-marketing context skill and working with Claude Code, Codex, Cursor, Windsurf and any Agent Skills client. v2.11.25 attaches one ZIP per skill for uploading in Claude's web and desktop apps, and v2.11.26 on 8 October adds a guide to packaging owned courses or newsletters as installable skills. Paid partner integrations fund the work and are disclosed. MIT, 54,098 stars, up from about 11,000.

  • Agent-Reach

    Panniantong · Agent Web Access

    Moved from Week 32: a single CLI and agent skill that lets AI agents read and search Twitter/X, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu, LinkedIn and other platforms without paid APIs, routing each platform across several backends. Since August it has hardened its trust boundaries, hidden setup credentials, added Windows CI and added a Boss Zhipin job-search channel on 15 September. There has been no new tagged release since v1.5.0 in June. MIT, 95,545 stars, up from about 69,000.

  • flow-mcp

    GTAI-1 · Google Flow MCP

    An MCP server that drives Google Flow (covered here in week 27) from Claude Code, Claude Desktop, Codex or Antigravity, or from a local control panel, spending Flow's own credits instead of the pay-per-use video API; free Google accounts get 50 a day. It runs locally and controls a dedicated Chrome profile over the Chrome DevTools Protocol, clicking Flow's real buttons after a one-time manual sign-in, with no extension or hosted service. Long calls return a task id collected later with flow_wait, and optional FFmpeg steps join clips, mix music and record narration. Developed on macOS, Windows untested. No licence file, 67 stars. Automating Flow's interface may conflict with Google's terms.

Developer Tools (2)

  • Impeccable v4.5

    Paul Bakaus · Agent Design Skills

    A design skill set, at impeccable.style, that gives coding agents a design vocabulary through commands such as typeset, distill, clarify, polish, audit, layout, colorize and animate, plus 59 automated checks for common AI design defaults. It works with Claude Code, Codex CLI, Cursor, Copilot, Gemini CLI, OpenCode and others through npx, plugins or skills.sh. The 4.5 series, released 2 to 9 October, builds app screens and docs code-first, lets you pick from HTML sketches, and filters false findings on scanned pages. Apache-2.0, 79,479 stars.

  • LLM Wiki v2

    crow-hq · Self-Hosted LLM Wiki

    A local Python app that files dropped .txt, .md and .pdf documents into a wiki of plain Markdown notes in ~/llm-wiki, answers questions with clickable citations to the notes it used, and shows a live map of folders and links. By default it runs in CROW mode, where a small typed classifier decides where each document goes instead of the main model; the repo reports 89% right-note reads against 62% for the classic approach, 2.1x faster ingest and 68% fewer tokens, self-reported. Works with OpenRouter, OpenAI, Gemini, Ollama or any OpenAI-compatible server, with Docker and an HTTP API. AGPL-3.0, created 24 September and pushed 9 October.

Research & Safety (6)

  • OpenAI Math Manuscripts

    OpenAI · AI-Generated Mathematics

    A repository, published 6 October, of 719 mathematics manuscripts in 372 families written by an unreleased internal OpenAI model, which the company pointed at about 4,000 open research problems after its existing math evaluations saturated. Each result used on average about three hours of ChatGPT Pro thinking compute, and about 42% of headline results (300 of 719) come with Lean formalizations; the README warns that unformalized results may contain errors. Reasoning summaries are included for ten families, among them the irrationality exponent of pi and the Mezard-Parisi spin-glass formula. On 7 October OpenAI withdrew three Hodge-conjecture-related papers over a sign error and revised 14 others. Most results are not peer reviewed. Apache-2.0, 13,745 stars.

  • AlphaProtein Novo

    Google DeepMind · De Novo Enzyme Design

    Google DeepMind's open-sourced pipeline for designing new enzymes, announced 5 October with a bioRxiv preprint written with Frances Arnold's lab at Caltech and Peng Liu's group at Pittsburgh. A diffusion model designs protein structure and amino-acid sequence together around a given catalytic motif and ligand, optionally followed by LigandMPNN sequence redesign and AlphaFold 3 structure filtering. The preprint reports more than 5,600 designs tested across five reactions with hit rates up to 80%, nitrene transferases that make piperidine at 99:1 regioselectivity, and an esterase that breaks down the plasticizer DEHP and is 14 times more active at 90°C than at room temperature. Code is Apache-2.0; model weights are a separate non-commercial download.

  • Context Language Models

    University of Washington and Meta Superintelligence Labs · Self-Managed Context

    Language models that manage their own context by treating it as a file they can rewrite, an idea that extends to multi-agent systems where each agent's context is a file. Built zero-shot from existing models, CLMs report 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus and 5% higher scores with 59% fewer FLOPs over a 12-hour EdgeBench run. Online reinforcement learning lifts Qwen3.5-9B by 47.6% on BrowseComp-Plus while using 12% fewer FLOPs, and a co-designed Suffix Cache Reuse cuts serving compute by 35% against SGLang, all self-reported. Posted 29 September; code, a Harbor harness and training modules are at facebookresearch/context-language-models, still active this week.

  • Living Models: Gemma 4 and BOTANIC-1

    Living Models / Google DeepMind · AI For Plant Genomics

    A Google Gemmaverse case study published 6 October. Living Models, an AI-biology lab in Paris, pairs Gemma 4 E4B with BOTANIC-1, its foundation model pretrained on DNA from 320 plant species, to find the mutations behind crop traits. Gemma 4 acts as the agent: it writes code, runs bcftools and Ensembl VEP, and sends plain-language questions to BOTANIC-1. In a melon test the system ranked the known causal mutation first out of 2,494 candidates in under four minutes, and across 20 runs it found it first 90% of the time, against 15% for expert-guided classical analysis, vendor-reported. The workflow runs locally on one NVIDIA L4 GPU through Ollama, and BOTANIC-1 weights (September) are on Hugging Face.

  • Vals: AI-Designed Magnetic Semiconductors

    Vals AI · AI Materials Discovery

    Vals AI, covered here for its Finance Agent v2 benchmark (week 23) and Vals Index (week 30), had Claude Opus 5.5 agents run density functional theory simulations at the PBE+U and HSE06 levels in search of antiferromagnetic semiconductors for future computer memory. The agents designed a new compound, YBaMnFeO5, with a predicted 2.35 eV band gap and magnetic ordering up to about 490 K, and resurfaced KV[Cr(CN)6], a 1999 material with a 2.1 eV gap and measured ordering up to 376 K. YBaMnFeO5 is a simulated prediction that has not been made in a lab. Inputs, raw outputs, analysis code and a caveats list are on GitHub. Published 4 October.

  • MotionSpaceFlow

    LY Corporation · Text-to-Motion Research

    A flow-matching method from LY Corporation for text-driven human motion generation that works on full-resolution motion instead of a learned, downsampled latent space, with no motion encoder or decoder. It adds representation-aware noise scaling and a multimodal diffusion transformer that uses causal attention for frame-to-frame motion features and bidirectional attention for absolute joint coordinates. The authors report state-of-the-art text-to-motion results on HumanML3D and other datasets, self-reported, and the XYZ variant supports zero-shot control of any joint at any frame without control-specific training. Paper posted 28 September; code and pretrained models followed on 8 October under CC0, in a repo the authors call a temporary release.

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