Quanta AI Labs

39 / Sep 14 – Sep 20, 2026 · 30 RELEASES

AI News Week 39, 2026

Jev's Moat Lasted Five Days. TypeSafe launched Jev on 15 September as a new class of model, a non-generative decision engine returning typed answers with calibrated probabilities instead of text, priced at $0.042 per million input tokens with output free. On 19 September Convai Innovations open-sourced Laya, which does the same job under Apache 2.0 and is free to self-host, reporting 32.8ms single-GPU latency against Jev's 236-276ms and better calibration at 0.081 expected calibration error against 0.246. Those are Convai's own numbers and nobody has re-run them, but the speed of the response is the story: a week after a $40M seed announced a new category, a competitor shipped the category as open weights. The other thread is AI building its own infrastructure, from two directions at once. Z.ai published how GLM-5.3 acted as an infra agent to build and optimise the production inference stack now serving GLM-5.3-Flash across more than 100,000 Chinese-made accelerators, going from first successful run to production in under two weeks and roughly tripling throughput. Google, DeepMind and two universities released Dream-RSI, which has a coding agent build a discovery tree, turn that tree into a simulator, then evaluate thousands of candidate policies offline at zero executions before redeploying, reporting 162x fewer agent calls than its baseline on Lasso tasks. Elsewhere Qwen shipped omni-modal and live-translation models in the same week, Xiaomi put its reinforcement learning runs on a public dashboard while the models are still training, and PrismML squeezed Qwen3.8 27B to 1.76 bits per weight and 5.9GB while keeping 98.2% of full-precision scores.

Language Models (6)

  • Jev

    TypeSafe AI · System One Model

    A model that stops generating text. You send unstructured input plus named questions and it answers all of them in one parallel pass, returning a choice from your own option list, a position on a scale you define, or a probability that a statement is true, each with a confidence figure. Choice caps at 255 options, input is capped around 32,000 tokens, pricing is $0.042 per million input tokens with output free. The advertised 0% hallucination rate means zero schema violations rather than zero wrong answers, and TypeSafe's own documentation says calibration "does not guarantee that an individual answer is correct". Their benchmark scores Jev against the average of GPT-6 Astra and Fable 5.1 outputs, which measures agreement rather than correctness, and the CEO conceded on Hacker News that the latency comparison is problematic because output tokens are not comparable. Launched 15 September with $40M led by DCVC.

  • Laya

    Convai Innovations · Open Decision Engine

    Released four days after Jev and doing the same job: a non-autoregressive decision engine running choice, score and boolean primitives over text and JSON rather than generating free text. Convai reports 32.8ms single-GPU latency, 7.2ms per question when batched, against Jev at 236-276ms, plus better calibration at 0.081 expected calibration error versus 0.246 and higher accuracy on AG News at 0.950 against 0.910. Over 100 languages with routing across non-Latin scripts. Apache 2.0 on Hugging Face, PyPI and GitHub, free to self-host. Every comparison is Convai's own and none has been independently re-run.

  • Ternary Bonsai 2 27B

    PrismML · Ternary Quantisation

    Qwen3.8 27B quantised to ternary weights, meaning each weight is -1, 0 or +1 with FP16 group-wise scaling, landing at 1.76 effective bits per weight and a 5.9GB footprint, over nine times smaller than full precision. It keeps the 262K context and multimodal text-and-image input. PrismML's own comparison against full-precision Qwen3.8 27B puts the aggregate at 83.9 against 85.4, or 98.2% of full-precision performance, with 81.58 against 82.17 on coding, 96.57 against 97.06 on maths, 78.59 against 81.64 on vision and 77.57 against 79.74 on agentic tasks. Apache 2.0 with weights out now, running through custom low-bit kernels on NVIDIA GPUs and Apple devices.

  • Occamy-1.0

    Accio Lab · Long Horizon Agent

    A 35B mixture-of-experts with 3B active per token over 40 layers and 256 experts, eight routed plus one shared, with a vision encoder, built on the Qwen3.6-35B-A3B checkpoint through full-parameter supervised finetuning, HDPO, model merging and SAO post-training. Training used 14,998 trajectories totalling 403.3M tokens across general agentic work, long-horizon interactive agents, terminal and software engineering, and tool-call grounding, at a 262,144-token architecture context. Accio reports Claw-Eval average rising from 69.50 on the base model to 82.20, with Pass-cubed at 71.40, ahead of GPT-5.6 Sol at 68.90 but behind Qwen3.8-Max at 73.68. Apache 2.0 with open weights, and the Dressage training framework is open too.

  • ZGCM-1

    Zhongguancun Academy · Math Reasoning Model

    A 7.39B dense model trained from scratch with a 256K context and both thinking and direct-response modes, aimed at mathematical reasoning and tool-assisted search. The repository is unusually complete for a release of this size: the full pretraining, mid-training, supervised finetuning and reinforcement learning pipeline with per-stage documentation, an arXiv technical report, architecture specs and results across more than 20 evaluations. Reported figures include 97.13% on MATH-500, 75% on AIME 2026 and 63.09% on WebWalkerQA, with a hybrid-attention design claimed to give 3.94x training throughput at 256K context against full attention. MIT, with weights and training data on Hugging Face.

  • Bespoke-Nimble-9B

    Bespoke Labs · Structured Scoring Adapter

    A LoRA adapter of roughly 165MiB on top of Qwen3.5-9B rather than a standalone model, which the name rather hides, so the base checkpoint is needed separately. It is specialised for context-grounded structured prediction: booleans, multiple choice up to 26 options and rubric score levels, scoring the allowed answer tokens directly instead of generating free text, and returning probability-weighted expected scores for rubric fields. Prompts over 2,048 tokens are rejected rather than truncated, and it needs a CUDA GPU with BF16. Apache 2.0 on Hugging Face, with no benchmark comparisons published.

Vision & Image (1)

  • LightOnOCR-2-1B

    LightOn AI · Document OCR

    Moved from Week 4 2026, and unchanged since: the model card and its arXiv paper are both dated 19-20 January 2026 with no revision or changelog since, so this is the same release rather than a new one. A 1B end-to-end vision-language OCR model under Apache 2.0 turning PDFs, scans and images into clean ordered text without a conventional OCR pipeline, trained with stronger French, arXiv and scan coverage than the original plus better LaTeX handling. Reports 83.2% on OlmOCR-Bench overall at 5.71 pages a second on an H100, 3.3x faster than Chandra OCR, for under a cent per thousand pages. Six variants released.

Video & Animation (3)

  • Odyssey-3

    Odyssey · World Model

    Announced 15 September as Odyssey's most capable foundation world model, an autoregressive diffusion transformer trained on a large corpus of visual observations to build a working understanding of physics, dynamics, cause and effect, and human behaviour. Named uses run from controlling robots and humanoids to driving vehicles, piloting drones, training other models and playing games. The data-efficiency claims are the substance: usable robot control from tens of hours of demonstrations and autonomous driving from 20 hours of simulated data. The announcement publishes no resolution, frame rate, latency or interactivity detail, and release is only given as coming weeks.

  • JING and DAO

    XGEN Labs · Egocentric World Model

    A two-part world simulation. JING is an egocentric interactive experience model generating first-person video and audio conditioned on actions, reference images and observation history; DAO is a separate shared-world engine holding persistent state and rules and computing what each observer should see as a consequence. Only JING has shipped, built on MiniMax-H3's Ref2VA with FlashGen for four-step bidirectional inference, open on GitHub and Hugging Face under the MiniMax H3 Community Licence with a JING-Flash-v1 checkpoint. The causal model and technical report are both listed as coming, and DAO has no code or weights. No parameter count, benchmarks or resolution published.

  • Meridian

    Viggle AI · Novel-View Video

    Re-renders an existing video from new camera angles and timings without a reshoot: orbit, pan, zoom, vertical moves, slow motion, pause and speed-up, and it can invent a camera move from a single still. Built on MiniMax-H3's transformer and VAE with no text encoder at inference, shipped as two 2.5GiB LoRA adapters rather than a checkpoint, using VGGT-Omega for depth and camera pose. Output is 1344x768 for 16:9 at 24fps, 73 to 243 frames, taking roughly 36 seconds at 73 frames and 150 seconds at 243 on a B200, at up to 113GiB peak memory. Weights under the MiniMax H3 Community Licence with territorial restrictions; code Apache 2.0.

Audio & Speech (6)

  • Gemini 3.8 Live

    Google · Live Voice Agent

    A real-time conversational voice model with near real-time visual grounding across 97 languages, including switching language mid-conversation, plus an Extended Thinking variant that reasons over multiple steps while still speaking, filling the gap with verbal acknowledgements. Google reports an Artificial Analysis Speech-to-Speech Index of 82.6 for Extended Thinking, 68.6% on tau-Voice agentic task completion, 35.1% on Sierra tau-Voice-banking and 97.7% on Big Bench Audio reasoning, with standard 3.8 Live placing second on Speech Agent Arena. Both can call tools mid-conversation. Live now in the Gemini API and AI Studio; Extended Thinking is limited to Pro and Ultra subscribers with enterprise access in private preview.

  • Qwen3.8-LiveTranslate

    Alibaba Qwen · Live Speech Translation

    Simultaneous interpretation built on an interleave architecture with a hybrid mixture-of-experts backbone split into Thinker and Talker modules. It recognises 60 source languages and speaks 29 of them, does per-speaker diarisation with individual voice cloning in real time, and cuts average lagging from 2.8 seconds in the Qwen3.5 predecessor to 2.3. Released 19 September through Alibaba Model Studio with a trial endpoint. Note that qwen.ai renders as an empty JavaScript shell to automated fetches, so these figures come from Qwen's own announcement and secondary coverage rather than the page itself.

  • YuE2

    m-a-p · Open Music Weights

    Moved from Week 37 2026, where it was covered from the project page. The weights are now downloadable: lyrics and a style prompt become a full song with vocals and accompaniment through an AR-NAR mixture-of-transformers backbone writing score and semantic tokens, with flow matching for the acoustic latents, plus cover-song generation via melody transcription. Runs locally on a 24GB GPU in BF16 without quantisation, licensed CC BY-NC 4.0 so non-commercial only. On WildSongBench it reports 6.9632 best-of-8 against Suno v5 at 6.8721 and v6 at 6.5562. One caution: the repo is named YuE2-3B but its own tensor spec says 4B, so treat the parameter count as unsettled.

  • R2T2

    NetEase Youdao · Streaming ASR

    Open-source streaming speech recognition for real-time transcription in English and Mandarin, with code on GitHub and weights on Hugging Face and ModelScope alongside a hosted demo. The site runs side-by-side transcription comparisons against a named competitor, GPT-Live-Transcribe, across banking, environmental policy and retail audio. A full technical report is described as forthcoming, and no pricing is listed.

  • Pocket TTS Training Guide

    Kyutai Labs · TTS Training Guide

    A training guide inside the pocket-tts repository explaining how to train a custom model on your own GPU and speech data, aimed at getting community models for new languages. It covers requirements at roughly 60GB of disk per 1,000 hours on Linux with an NVIDIA GPU, data preparation from HiFiTTS-2 or your own sources, forced alignment, training from scratch or by distillation or finetuning, evaluation on WER, speaker similarity and UTMOS, H100 benchmark figures, and troubleshooting for transcript and alignment problems. Pocket TTS itself is a TTS built on the CALM architecture that runs on CPU. MIT, 9,572 stars.

  • OpenLess v2.0.0-Beta.1

    Open-Less · Voice To Prompt

    A cross-platform voice input tool, Rust core with Tauri on macOS, Windows and Android and egui on Linux, that transcribes speech held under a hotkey, polishes it with an AI provider and inserts the result at the cursor, aimed at turning dictation into structured prompts. It has a style-pack marketplace, a self-learning correction dictionary and bring-your-own ASR and LLM credentials for local-first use. 3,595 stars. Worth flagging for anyone who adopted it earlier: the licence changed from MIT in the 1.x line to AGPL-3.0-only as of this 2.0.0-Beta.1 release on 10 September.

3D & Spatial (1)

  • PlayCanvas Engine v2.22.2

    PlayCanvas · Engine Patch Release

    Since Week 19 2026, the engine has shipped three point releases: v2.22.0 on 4 September, v2.22.1 on 8 September and v2.22.2 on 11 September, the last fixing shadow shader selection when cascade count changes and a StandardMaterial texture-map iteration bug. All patch level rather than a major version. The engine itself is an open-source WebGL and WebGPU game engine in TypeScript, MIT licensed, 16.8k stars.

Agents & Automation (7)

  • Qwen3.8-Omni-Flash

    Alibaba Qwen · Omni Modal Agent

    Qwen's first omni-modal model built around agentic use, taking text, image, audio and video across a 1M-token context while holding text-only performance level with a same-size text model. Qwen claims an average improvement of more than 25% over Qwen3.5-Omni-Plus across 29 evaluations, alongside API price cuts of over 98% for audio input and over 93% for audio-visual input. Generally available through Alibaba Cloud Model Studio on Beijing and Singapore endpoints. As with the LiveTranslate post, the qwen.ai page returned an empty shell, so the numbers are vendor-reported through secondary sources.

  • Astra for Law

    OpenAI · Legal Research Agent

    Launched 17 September in the United States, pairing GPT-6 Astra with a dedicated legal search index covering US case law, statutes, regulations, court rules and administrative decisions across more than 230 million sources, built with the Free Law Project and CourtListener and updated daily, plus legal-specific instructions and tooling. Reported gains are 40% more accurate research answers, 24% more reference cases found and 54% more relevant passages retrieved against Astra using plain web search. Access is through a Trusted Access programme in ChatGPT and Codex with the API to follow, with Latham & Watkins, Ropes & Gray, Cooley and Sullivan & Cromwell named as early adopters and 26 partner plugins. openai.com blocks automated fetches, so this rests on secondary coverage.

  • MiniMax Code

    MiniMax AI · Terminal Coding Agent

    A terminal coding agent rather than a model: a CLI with a desktop wrapper that reads, edits and diffs code, runs shell commands and tests under permission controls, and handles session resumption, task planning and headless execution for CI. It works against MiniMax's own API or any OpenAI- or Anthropic-compatible provider with your own key, takes plugins and skills, and supports the Agent Client Protocol for editor integration. MIT for the first-party code, needs Node 22.19 or newer, 1.4k stars. No benchmarks published.

  • GPT-Policy

    Cheng, Yi, Fang and 12 co-authors · In-Context Robot Learning

    In-context robot learning with a fixed general-purpose vision-language model as the acting agent, named on the page as GPT-6 Astra, with no gradient updates or parameter changes between tasks. A context compiler feeds the frozen model, which proposes robot-tool actions, and a constrained controller validates and executes them, drawing on human video demonstrations, robot demonstrations with actions, target images, self-interaction history and human-robot interaction. Across 10 tasks in 5 experiment families it reports 2 of 3 successes on Pick Red Towel with human video against 0 of 3 without, and 3 of 3 on Unscrew Bottle Cap with richer context. Worth noting the sample size: three trials per condition, which the page does not caveat. Code published, licence not stated, institutional affiliation not given.

  • Pexo Skills

    pexoai · Video Agent Skills

    A collection of agent skills that lets hosts such as Claude Code, Cursor and OpenClaw turn text, images, URLs, scripts or audio into finished video with multi-shot sequencing, music and subtitles, taking 8 to 20 minutes depending on length. Rather than wrapping one model it selects among more than ten platforms per job, naming Seedance 2, Kling 3.0 and HappyHorse among them. Follows the Agent Skills Specification with a primary pexo-agent skill plus scenario, creative and studio sub-skills for image, video and audio. MIT, 785 stars.

  • MiniMax Design

    MiniMax · Multi-Agent Studio

    A downloadable desktop app for macOS 13+ and Windows 10+, labelled V.1.0 Director's Build, that runs several specialised agents in parallel from one creative brief, covering copy, image, video and audio. The workflow is five steps: idea intake, a canvas flow, reusable skills, local asset management, then review and delivery, all on a single canvas spanning scripting, storyboarding, assembly and editing. Aimed at short-form video, e-commerce content, brand marketing, short drama, commercials, audiobooks and podcasts. No pricing is shown despite a pricing link, and there is a creator revenue-share programme.

  • Navana.ai raises ₹40 CR Series A

    Navana Tech India · Indian Voice AI Funding

    Announced 25 September: an Indian voice-AI infrastructure company raising a ₹40 crore (roughly $4.8M) Series A. Its stack is BODHI, proprietary speech models covering 12+ Indian languages and 40+ dialects, an end-to-end telephony voice-agent platform, an API layer, and batch audio intelligence for transcripts and summaries. States RBI-aligned compliance, ISO 27001 and SOC 2 Type II, and cites 1.5 million users on its Hello Ujjivan deployment with named customer Ujjivan. No model architecture, parameter count or benchmark figures published; the page is otherwise an evergreen company homepage built around this one dated funding announcement.

Developer Tools (4)

  • Toward Recursive Self-Improvement

    Z.ai · Self-Built Inference Stack

    Z.ai describes putting GLM-5.3 to work as an infra agent that built and optimised the production inference stack now serving GLM-5.3-Flash on a cluster of more than 100,000 Chinese-made accelerators. From first successful run to production readiness took under two weeks, with end-to-end throughput roughly tripling against the initial baseline and per-token cost said to reach parity with mainstream Nvidia GPUs. The obstacles listed are the interesting part: limited chip memory and bandwidth, a new model architecture, a 1M context window, multimodal requests, an immature software ecosystem and incomplete kernel support. Published 17 September; z.ai fetches empty, so sourced from secondary coverage.

  • Needle 3

    Cactus Compute · Tiny Device Model

    Since Week 33 2026, when this archive covered Needle 2. Needle 3 is an automation model for tiny devices at 29 to 121M parameters using laddered simple attention networks with CQ2 quantisation, packaged at 8 to 29MB for mobile, wearables, robots, smart home, automotive and microcontrollers. The new mechanism is intelligence laddering: one set of weights yields a usable model at any depth from 2 to 20 layers. Cactus claims a 4-layer tuned version passes DeepSeek V4 Flash on structured tasks and beats models ten times its size on mobile tool calls, at 400 to 4,000 tokens a second decode on a Raspberry Pi 5. No licence or release date stated on the page.

  • God's Eye View

    Bilawal Sidhu · Live 3D Globe

    An open-source browser globe that puts live aircraft, ships, satellites, earthquakes, traffic and public cameras on photorealistic 3D terrain, described by its author as a spy satellite simulator except the data is real. Built on CesiumJS and Vite, aggregating open feeds from OpenSky, adsb.lol, AISStream, CelesTrak, USGS, NASA FIRMS and OpenStreetMap, with a voice-controlled agent on the OpenAI Realtime API. MIT, 38.8k stars and 7.8k forks, created in June 2026 and still being pushed to in September. No tagged releases.

  • Inspo

    Nutlope · Design Reference MCP

    An MCP server giving coding agents a reference archive of real production websites to look at before building UI. It ships 2,320 captured pages across 832 sites with extracted design systems covering palettes, type ramps, spacing and CSS variables, plus 68 canonical reference components with JSX examples, exposed through 15 tools for semantic search, similarity search and component discovery, on a read-only SSRF-protected architecture. TypeScript MCP server over stdio or HTTP with a Next.js 16 and Tailwind v4 gallery, Playwright for captures and Together AI for embeddings. MIT, 606 stars.

Research & Safety (2)

  • Dream-RSI

    Google, DeepMind, Maryland & Virginia · Recursive Self-Improvement

    Recursive self-improvement through evolving worlds. A coding agent explores and builds a discovery tree, that tree is then reused as a simulator, and thousands of candidate policies get evaluated against the simulator at zero executions before the best one is redeployed. Reported results include 2.43x fewer generations on VGG16 at comparable performance, 2.09x higher scores on ConvDiv at comparable budget, and 162x fewer discovery-agent calls than the baseline SimpleTES on Lasso tasks, with experiments run on Gemini 3.1-Pro and 3.7-Flash across algorithm engineering, mathematical optimisation and GPU kernel work. Code released.

  • Hacking OpenAI

    Hacktron · Disclosed Exploit Chain

    A responsible-disclosure write-up chaining a heap buffer overflow in libheif, reached through Discourse's ImageMagick-based HEIF decoder, with an SSO misconfiguration in OpenAI's identity infrastructure. The researchers report that a malicious HEIF image uploaded to the community forum gave remote code execution, which they then escalated through SSO from forum accounts to ChatGPT and Codex access. Reported to OpenAI through Bugcrowd and to Discourse through HackerOne on 25 July 2026, with OpenAI confirming a fix the same day and paying a $6,500 bounty, and Discourse publishing advisory GHSA-vhm9-85gw-x335 on 28 August. OpenAI noted that testing against the Discourse-hosted forum was outside bounty scope.

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