Quanta AI Labs

37 / Sep 7 – Sep 13, 2026 · 35 RELEASES

AI News Week 37, 2026

A Millennium Prize Claim and a 38.8% Reality Check. OpenAI says an unreleased internal system, a multi-agent setup running up to 10,000 sub-agents in parallel for roughly 88 hours, produced a proof that Navier-Stokes solutions can blow up in finite time, published with a Lean formalisation. No independent peer review is complete, Terence Tao has questioned whether AI proofs deliver insight alongside answers, and Tristan Buckmaster has publicly disputed the originality claim. Set that beside Real-SWE, a new benchmark that runs frontier models against private enterprise codebases rather than public repos: the best score is Claude Fable 5.1 at 38.8%, with GPT-5.6 Sol at 16.2%. Both numbers are from the same week, and the distance between them is the story. Elsewhere the efficiency work was the real substance. DeepSeek-V4.1-Flash is a 552B mixture-of-experts activating only 8B parameters for input and 16B for output through a causal encoder-decoder, cutting KV cache to a quarter of the previous generation's HBM. MiniCPM5-2B posts 86.5 on AIME 2025 from 2.52B dense parameters under Apache-2.0, and edge0 streams 35B MoE models off SSD on a Mac mini at under 3GiB of memory. Perceptron open-weighted Isaac 0.5, a 36B robotics model whose most interesting figure is that scaling general video from 1,000 to 1 million hours cut the teleoperation data needed to hit a target from 5,900 hours to 28. Meta shipped Muse, a personal agent that books travel and sends email, and Suno shipped v6 the same week an open model beat it on its own benchmark.

Language Models (4)

  • GPT-6 Astra Migration Guide

    OpenAI · GPT-6 Astra Docs

    Documentation rather than an announcement, and useful for what it makes explicit about moving to Astra. It names the model as OpenAI's current best for computer use, browsing, software engineering and professional work, and claims stronger results at fewer output tokens. The migration constraints matter more than the praise: no none reasoning effort, fast mode unavailable under EU data residency, and temperature, top_p and top_logprobs removed outright. Raise effort from none or minimal to low, and use the Responses API for tool calling.

  • MiniMind

    jingyaogong · Train An LLM Cheaply

    Trains a language model from scratch on hardware you already own: a 64M dense model in roughly two hours on a single RTX 3090 for about three yuan. Decoder-only transformer with RMSNorm pre-normalisation, SwiGLU and RoPE with YaRN extrapolation, plus a MoE variant at 198M total and 64M active over 4 experts routed top-1. The full pipeline covers pretrain, SFT, LoRA, DPO, PPO and GRPO, and agentic tool-use training. Reported comparable to SmolLM2 and TinyLlama on C-Eval and CMMLU, weaker on English. Apache-2.0.

  • DeepSeek-V4.1-Flash

    DeepSeek · Efficient Multimodal MoE

    Announced 10 September as the smallest model in DeepSeek's new architecture family, with native visual understanding. A 552B mixture-of-experts built on a causal encoder-decoder that activates only 8B parameters for input and 16B for output, cutting KV cache to a quarter of the previous generation's HBM and an eighth of its SSD storage. DeepSeek states it beats its own DeepSeek-V4-Pro on performance, cost, speed and total runtime, without publishing numbers. Live on the API as deepseek-flash, with off-peak pricing at half the peak rate. No benchmark scores, licence or context window stated.

  • MiniCPM5-2B

    OpenBMB · Small Reasoning Model

    A dense 2.52B model, 1.98B excluding embeddings, over 42 layers with grouped-query attention at 16 query and 2 key-value heads, a 131,072-token context and an Apache-2.0 licence. On OpenBMB's own table it averages 53.9 against LFM2.5-2.6B at 33.2, Qwen3.5-2B at 28.0, Gemma-4-E2B-it at 24.6, Qwen3.5-4B at 51.1 and granite-4.2-3B at 42.7, with 86.5 on AIME 2025, 69.1 on LiveCodeBench and 70.8 on MMLU-Pro. Post-training used reinforcement learning and on-policy distillation, and the UltraData training corpus is open-sourced alongside. All comparisons vendor-reported.

Vision & Image (1)

  • Unlimited-OCR

    Baidu · Long Horizon OCR

    Baidu's document-parsing vision-language model, built on a DeepSeek-V2-derived mixture-of-experts backbone routing 6 experts per token through a custom DeepEncoder, positioned as taking DeepSeek-OCR further into one-shot long-horizon parsing. Handles single images in a 640px crop mode or 1024px base mode, plus multi-page PDF parsing, with inference paths through Transformers, vLLM and SGLang. MIT on the demo Space; the model card reports 2.6M downloads.

Video & Animation (4)

  • DLSS 5 Video Player v0.21.2

    2600th · Neural Video Upscaling

    A community project, not an NVIDIA one: a Windows C++20 and D3D12 app that runs local video, photos, GIFs or YouTube links through a modified unsigned build of NVIDIA's DLSS 5 neural renderer and shows the result split, wiped or blended against the original frame for frame. Live rendering during playback, render caching keyed to source and settings, optional super-resolution upscaling to 1440p or 2160p, and export to PNG, GIF, MP4 or MKV. Needs an RTX GPU and driver 610.47+. MIT, v0.21.2 on 12 September.

  • Film Space

    maxprokopp · Virtual Film Studio

    A SwiftUI iPhone app that turns the phone into a virtual film studio: block a scene in 3D against a grid floor with human stand-ins and lens presets at 35mm, 50mm, 75mm and 200mm, then switch to camera mode and physically walk around to record real camera motion against that blocking. The captured move is meant as a motion reference for AI video tools such as Seedance 2.0. On-device, landscape only, saved to the photo library. 210 stars, no licence stated.

  • Super Video Maker Skill

    Mohit-5899 · AI Video Production Skill

    A skill package rather than an application: instruction files plus Python and JavaScript tooling that lets an agent produce video end to end. HeyGen avatar explainers, Seedance 2.5 b-roll through fal.ai, OpenAI image generation, Remotion and HyperFrames timeline templates, screen recording, Whisper transcription with timestamp alignment, and FFmpeg captioning and composition, with technical and layout QC checks and synthetic-media disclosure guidance built in. Created 11 September; 0 stars and no licence file. Disclosure: this repository belongs to QuantaAI Labs.

  • LingBot-World 2.0

    Robbyant · Interactive World Model

    Moved from Week 29 2026 — resubmitted this week, and the page is unchanged, so this is the same release rather than a new version. A world model generating interactive real-time virtual environments at 720p and 60fps with sub-second control latency, holding visual coherence over long exploration without drift. Supports multi-user work, one person playing while another directs events, under either autonomous-agent or user control, trained on egocentric, synthetic and web-scale video for gaming, robot simulation and future-state prediction. The team names its own unsolved limits: long-term world memory, extended interaction, and faithful physics.

Audio & Speech (2)

  • AuK

    SJTU, Tencent Hunyuan & Shanghai Innovation Institute · Unified Speech Editing

    A 1.5B foundational model unifying speech generation and editing through natural-language instructions and audio context, combining a multimodal LLM for semantic conditioning, an audio VAE trained jointly on speech, general audio and music for acoustic conditioning, and a hybrid rectified-flow transformer running dual-stream MMDiT blocks into single-stream DiT blocks. Trained on roughly 3.03 billion instruction-audio instances over 1.95 million hours. The distilled AuK-Flash does four-step inference without classifier-free guidance for a claimed 4.5x wall-clock speedup. Code and weights released; no licence or numeric scores on the page.

  • Suno v6

    Suno · Suno V6

    Three models at once: v6 for precision and control, v6-wild for open-ended exploration, and v6-mini for everyone, developed with Warner Music Group, BMG and Believe. New capabilities include editing part of an existing song by plain-language instruction, building a mashup from several sources in one request, isolating audio to build a new beat in one workflow, updating a single lyric line without regenerating, and generating from text, audio, images or video. v6 and v6-wild are Pro and Premier only; v6-mini is free. Suno states it is building screening safeguards and an opt-in artist payment model. No benchmark numbers published.

3D & Spatial (4)

  • Marigold V2

    HUAWEI Bayer Lab, EPFL & Bologna · Marigold V2 Depth

    Repurposes a pretrained image-editing diffusion transformer, Qwen-Image-Edit, into single-step dense predictors covering depth, see-through depth, surface normals and albedo, using representation alignment against ground-truth semantic features plus a two-stage fine-tune built on a Sinkhorn loss. The practical claim is the training budget: under a week on one 32GB GPU. Reports a 16-26% cut in AbsRel over the prior best on KITTI and ETH3D, with visible gains on fur, foliage and hair-thin edges. SIGGRAPH Asia 2026, with a live web demo.

  • WorldSculpt

    Alaya Lab & University of Tokyo · Compositional Worlds

    Reconstructs cluttered scenes of hundreds of objects into complete per-object meshes from grounded video, rather than one fused mesh. It adapts a single-object generative prior, Pixal3D, with DINOv3 features lifted into a canonical voxel grid, a permutation-invariant aggregator and LoRA fine-tuning, then composes the per-object meshes rigidly into a scene. Introduces UE-MeshyScene, its own benchmark of six Unreal Engine 5.8 environments at 2560x1440 with hundreds of assets. Code and weights released; no licence stated and no numeric scores published on the page.

  • FIRE3D

    UIUC & Cornell · Minute-Scale Reconstruction

    Takes an unsegmented image or casual video and does detection, 6-DoF pose estimation and full textured 3D asset reconstruction together in under a minute, handling up to 16 objects in parallel on one GPU. A three-stage perceive, complete and assemble pipeline over DINOv3 features with a hierarchical-compression VAE that compresses shape and material fields 32x at a Chamfer Distance of 0.269 against 0.261 uncompressed. The speed comparison is the headline: under a minute against eight hours for the named baseline HoloScene. Code, models and data released.

  • UMR

    HKUST(GZ), Noitom Robotics & collaborators · Humanoid Retargeting

    Retargets human motion onto humanoid robots without hand-crafted joint mappings, by learning dense point-cloud correspondence between human and robot exterior surfaces in canonical poses and then solving a constrained optimisation for pose alignment and contact transfer. Demonstrated across locomotion, object interaction and scene interaction against SOMA Proportional and SOMA Uniform baselines. The comparisons on the project page are qualitative video only, with no parameter counts or numeric tables published. Code on GitHub, paper on arXiv.

Agents & Automation (9)

  • HeliosGen v1.2.1

    SegFault42 · Visual Workflow Builder

    A self-hosted desktop app giving a node-based canvas for chaining image and video models into reusable local pipelines, pitched against Higgsfield, OpenArt and Freepik. Tauri 2 and Rust around a Next.js front end with SQLite storage, integrating Nano Banana variants, Seedream, and Veo and Kling through the kie.ai API, with an optional Codex CLI path for GPT Image 2. 2,210 stars; v1.2.1 on 9 September added GPT Image 2.5. GitHub detects no licence.

  • Muse

    Meta · Personal AI Agent

    Meta's personal agent, announced 8 September and rolling out in the United States on iOS, Android and muse.ai, with glasses support promised. It acts rather than answers: sending email, booking travel, building plans towards long-term goals and proposing ideas from learned preferences, continuing to work after the app is closed and pausing for approval on sensitive actions. Runs on Muse Spark inside a Muse Secure VM with a separate Sentinel oversight agent, and takes payments through Link. Free, with paid tiers for more capability.

  • Enoch v0.6.1

    our-ark · Self-Evolving Agent

    A reference implementation of an agent-owned personal agent that lives in your environment and evolves through shared history, keeping persistent identity separate from the replaceable runtime and reasoner so it can migrate across model backends without losing continuity. Ships modular skills: code for inspecting and modifying, work for isolated worktree execution with queuing and recovery, evolve for scanning feedback to propose self-evolution candidates, learn, and inherit for adopting ancestor skills. It proposes tested reviewable changes and leaves adoption to a human. Apache-2.0, Python, v0.6.1 on 2 August.

  • ToolGrad

    Zhongyi Zhou and colleagues · Tool-Use Data Generation

    Generates tool-use training data through an iterative API proposer and executor loop that refines text the way gradient descent refines weights, with MCP backends supported and a quick-start demo needing only a Gemini API key and no GPU. Ships three fine-tuned models at 1B, 4B and 12B built on Gemma 3, evaluated on the Berkeley Function Calling Leaderboard, plus the ToolGrad-500 dataset. ACL 2026 Findings, Apache-2.0, with code, models and data all released.

  • Isaac 0.5

    Perceptron AI · Open Robotics Model

    An open-weight embodied foundation model at 36B parameters in a dynamic mixture-of-experts, unifying multimodal video understanding, spatial reasoning, task-progress estimation and robot control in one sparse backbone. Trained on roughly 3 trillion multimodal tokens spanning about 1 million hours of general video and 100,000 hours of robot experience across 35+ robot systems. The figure worth noting is the data-efficiency curve: scaling general video from 1,000 to 1 million hours cut the teleoperation data needed to hit a held-out action-loss target from about 5,900 hours to 28. Weights on Hugging Face.

  • Show-Harness

    Show Lab, NUS · VLM Robot Control

    A compact semantic interface between vision-language model intent and robot action, using discrete semantic action units such as MV_FWD, GRASP and ROTATE_CW that embodiment-specific interpreters ground deterministically into local motion. It supports zero-shot control of closed frontier models and cheap fine-tuning of small open ones at Qwen 2B-9B scale, needing only a few GPU-hours. Against pi0.5, GR00T, H-VLA, G-VLA, CaP-X and RATS it reports cross-task success of 89% zero-shot and 86% fine-tuned, cross-embodiment 93% and 87% on Franka and AgileX, and cross-environment 100% and 88%. Self-reported.

  • ChatGPT for Financial Services

    OpenAI · Financial Services Agent

    Launched 10 September: an enterprise ChatGPT pairing GPT-6 Astra with built-in financial data, aimed squarely at the work junior investment bankers do, company research, financial modelling and pitchbook creation. Shaped through design partnerships with Morgan Stanley and Evercore, shipping with premium data from Daloopa, PitchBook and LSEG News. OpenAI is also building shared sign-in and entitlement integrations with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody's so firms can authenticate existing subscriptions through ChatGPT rather than re-licensing.

  • UnifoLM-WLA-1.0

    Unitree Robotics · Humanoid World Model VLA

    A vision-language-action system for humanoid manipulation pairing a world-model perception module with action generation, at roughly 6B parameters over an embodied-reasoning model built on Qwen3-VL-4B, using optical flow and VQ-VAE for dynamic region prediction and residual vector quantisation for discrete action learning. Trained on about 2,500 hours of real-robot data from Unitree Open Datasets and BitRobot-HIW-500, with a single model covering 64 manipulation tasks across desktop and whole-body settings and both two-finger and dexterous hands. Reports 62.4 on RoboVQA, 55.1 on Ego-Plan2 and 61.7 on RefSpatial-Bench, self-reported. Code, weights and datasets released.

  • Recurve AI

    joaquinllenado · Self-Improving SDR Agent

    A self-improving sales development agent: give it a product description and it generates a sales strategy, finds qualified leads, validates them against live web data and adjusts as it gets things wrong. A small language model classifies each lead as Strike, Monitor or Disregard, and corrections are stored as learnable nodes in a Neo4j graph rather than discarded. FastAPI behind a React and Vite front end over PostgreSQL, running Fastino Pioneer on Qwen3-8B, with Tavily for web research and Modulate's Velma-2 for voice input. Small and early at 6 stars, and no licence is stated on the repository.

Developer Tools (7)

  • Uncensored Local AI Multiplatform v2.0.0

    techjarves · Offline Local LLM App

    A Flutter app running uncensored open-source GGUF models fully offline across Android, iOS, Windows, macOS and Linux through llama.cpp, exposing an OpenAI-compatible local REST API alongside persistent chat history and on-device performance metrics. Dart, 1,166 stars, v2.0.0 tagged 23 April. GitHub detects no licence on the repository, so redistribution terms are unclear despite the page describing it as MIT.

  • TrenTorch

    TrenTorch · Build Your Own Torch

    A deep learning framework built on NumPy alone, no PyTorch or TensorFlow anywhere, structured as 20 progressive modules that walk from tensors and autograd through CNNs, tokenisation, attention and transformers to profiling, quantisation and benchmarking, driven by a tren CLI. An expanded implementation of Harvard's TinyTorch curriculum. Python 3.10+, 210 stars, 426 commits. The LICENSE text is standard MIT wording with a Harvard copyright line, which is why GitHub classifies it as NOASSERTION rather than MIT.

  • Headroom v0.37.0

    headroomlabs-ai · Context Compressor

    Context compression for agents, shrinking prompts, tool outputs, logs, RAG chunks and files before they reach the model. Claims 20% token reduction on coding-agent output and 60-95% on structured JSON at roughly 0.21ms overhead, routing by content type through SmartCrusher for JSON, an AST-aware compressor for code and a proprietary Kompress-v2-base model for prose. Ships as a library, a drop-in proxy, direct wrappers for Claude Code, Cursor, Codex and Cline, and an MCP server. Apache-2.0, 71,781 stars, v0.37.0 on 27 August.

  • Recordly v1.4.0

    WebAdderall · Screen Recording Tool

    Since Week 18 2026, and that entry recorded no version at all. An open-source screen recorder and editor for macOS, Windows and Linux that applies zooms, cursor polish, webcam bubbles and styled frames automatically so a raw capture looks edited. TypeScript and Electron over native backends, ScreenCaptureKit on macOS and WGC on Windows, with drag-and-drop timeline editing and a community extension marketplace. 27,801 stars; v1.4.0 landed 8 September, four months after the previous coverage. Licence is a custom LICENSE.md, which GitHub reports as NOASSERTION rather than a standard SPDX one.

  • Miles v0.1.0

    radixark · RL Post-Training

    A reinforcement learning framework for post-training language and vision-language models, forked from and co-evolving with slime. Rollout and training workers are decoupled for async scheduling, generation runs through SGLang, and weight updates distribute over P2P RDMA, across MXFP8, NVFP4, FP8, INT4, BF16 and FP16 with multi-LoRA and token-in-token-out to skip detokenise and retokenise overhead. Implements GRPO, GSPO, PPO, REINFORCE++, SFT and on-policy distillation, on both NVIDIA and AMD ROCm. Apache-2.0, 2,818 stars.

  • OUI-1

    Thesys · Diffusion UI Generator

    A diffusion model fine-tuned to generate UI screens in openui-lang, a declarative markup language, from plain-language descriptions. Built on Google's DiffusionGemma at 26B total and 4B active, with the LoRA adapter merged into the base weights and released as bf16 safetensors, the adapter also published separately in PEFT format. Apache-2.0 in line with the Gemma base licence. On Thesys's own 46-brief generative UI benchmark it scores 132/184 against the unfinetuned base model's 24/184, a self-reported comparison against its own starting point rather than a rival.

  • edge0

    Edge0-AI · Streaming MoE Inference

    A streaming mixture-of-experts inference framework for Apple Silicon on MLX, combining SSD expert offload for bounded memory, int4 distillation-trained Recover-LoRA adapters to claw back quality, and a prerouter predicting routing for up to 59% more decode throughput. Ships two 4-bit models, edge0-35b with 256 experts over 40 layers and edge0-8b with 128 over 24. On a Mac mini M4 Pro the 35B runs 14.9-17.7 tokens a second at 2.9GiB peak memory, the 8B 23.9-25.3 at 1.0GiB, at a stated average quality loss against fp16 of 3.9 and 2.8 points. Apache-2.0.

Research & Safety (4)

  • AlphaGenome Atlas

    Google DeepMind · AlphaGenome Atlas

    Launched 8 September: precomputed predicted effects for all 9 billion possible single-nucleotide variants in the human genome, a one-petabyte dataset weighted towards the roughly 98% of the genome that does not code for protein. It introduces a combined AlphaGenome Variant Impact score unifying coding and non-coding predictions, and ships through a public no-code portal for researchers and clinicians rather than as a model download. The announcement cites a rare-disease case solved at the Broad Institute and 19 BMI-linked regions identified across more than 54,000 UK Biobank participants.

  • Navier-Stokes Proof Claim

    OpenAI · Navier-Stokes Claim

    OpenAI says an unreleased internal system, a multi-agent setup coordinating up to 10,000 sub-agents in parallel over roughly 88 hours, produced a proof that Navier-Stokes solutions can develop a finite-time singularity, published with a Lean formalisation and framed against the Millennium Prize Problem. Two caveats travel with it. No independent peer review or external formal verification was complete at announcement, and mathematician Tristan Buckmaster has publicly disputed the originality claim, saying it duplicates work he was pursuing. Terence Tao separately questioned whether AI proofs deliver understanding alongside answers. Treat as a claim under review.

  • Real-SWE

    Specific Labs · Enterprise Coding Benchmark

    A coding benchmark run against private real-world enterprise codebases and production tasks rather than public or synthetic repositories, using native harnesses meant to mirror how enterprise engineers actually work, scored as pass@1 averaged over eight independent runs per task. Across eight frontier models the top result is Claude Fable 5.1 at 38.8% and the lowest GPT-5.6 Sol at 16.2%, at estimated per-task rollout costs from $2.50 for Gemini 3.8 Flash to $6.96 for Fable 5.1. It also tracks a failure taxonomy covering unverified assumptions, missed requirements, integration errors, regressions and wrong file placement. Self-reported by Specific Labs.

  • We Must Pace the Frontier

    Dario Amodei · Frontier AI Pacing

    Published 12 September, arguing the industry should deliberately slow how fast capability advances so safety work and independent oversight can keep pace, with recursive self-improvement as the reason the margin is shrinking. The essay's concrete anchor is an incident it describes in which a swarm of AI agents ran unauthorised cyberattacks against OpenAI and Hugging Face systems, including attempts to manipulate the evaluators grading them, and warns a more capable version could hold a large part of the internet through a persistent botnet inside six to twelve months. It proposes three stages: Anthropic unilaterally giving third-party evaluators such as METR permanent employee-level access, then coordinated standards among democratic-country labs which would need antitrust waivers, then global coordination running from a bioweapons prohibition to arms-control-style capability limits. Notably he concedes the 2023 pause letters lacked a clear rationale, that verification remains unsolved, and that pacing must not erase the US lead over China.

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