Quanta AI Labs

40 / Sep 21 – Sep 27, 2026 · 23 RELEASES

AI News Week 40, 2026

Three Labs Cut Prices in the Same Week. Anthropic shipped Claude Opus 5.5 at 40% less cost and over 30% faster than Opus 5, xAI shipped Grok 4.7, and OpenAI shipped GPT-6 Sol and Luna at roughly half their predecessors' prices, all inside a few days of each other. Each vendor's own benchmark table shows gains over the other two by name, which is the more telling detail: frontier labs are now routinely publishing head-to-head comparisons against named competitors rather than only against their own prior release. The other running story is what happened to Jev, TypeSafe's decision-model launch from two weeks ago. Contrastive-LM's CLM-v0.1-8B benchmarks itself directly against Jev and claims 4-9x lower latency at comparable accuracy, joining Laya (week 38) as a second open-weight competitor to a model that was pitched as a new category five weeks ago. Xiaomi's MiMo-V2.6 finished what week 38 caught mid-training: both Pro and Flash have shipped as open weights, Pro a 1.02-trillion-parameter mixture-of-experts scoring 46.32 on the Artificial Analysis Intelligence Index, a third-party number rather than a self-reported one. Black Forest Labs moved FLUX from image generation into robotics with FLUX 3 Action, a world-action model predicting video and robot actions together. And Anthropic reports Claude agents found a genuinely novel enzyme system by scanning 200,000 candidates down to one, with the caveat, stated plainly by Anthropic itself, that nobody yet knows what the enzyme does.

Language Models (7)

  • GuppyLM

    arman-bd · Tiny Educational LLM

    An educational small language model at roughly 8.7M parameters over 6 layers, styled to talk like a small fish, built as a vanilla transformer with no advanced optimisations so the architecture stays readable. Trains from scratch in about 5 minutes on a single T4 GPU via Colab on 60,000 synthetic conversation samples, and runs locally via CLI or in-browser over WebAssembly. 3,810 stars. One page describing it as MIT-licensed does not match GitHub's own API, which shows no licence file at all, so treat the licence as unconfirmed rather than MIT.

  • MiMo-V2.6-Pro / Flash

    Xiaomi · Open-Weight MoE Models

    Since Week 38 2026, where these two models were only a live training dashboard with nothing shipped yet. Both have now actually shipped: MIT-licensed weights on Hugging Face as MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL, plus a MiMo-V2.6-Distill-Qwen-9B finetune. Pro is a 1.02T-parameter mixture-of-experts with 42B active and a 1M-token context, scoring 46.32 on the Artificial Analysis Intelligence Index v4.3, a third-party index rather than a self-reported number. Flash is 309B total with 15B active. Both are omnimodal, text, image, audio and video in, text out. Xiaomi states the final RL run cost roughly $2.62M for Pro and $850K for Flash.

  • Grok 4.7

    xAI · Frontier Reasoning Model

    Announced 21 September, using a larger base model than Grok 4.6 plus a longer reinforcement-learning run weighted toward tasks that take many hours, with native understanding of the Grok Bot interface and improved self-verification. Reports 46.3% on CursorBench 4.0 against 4.6's 40.4%, 71.0% on DeepSWE v1.1, 64.0% on the EEBench electrical-engineering benchmark, and only 3.3% of risky prompts passing through on HackerBench cyber-defence. $2 and $6 per million input and output tokens, with a faster variant at double that. Available through Cursor, Grok Build, the Grok API and third-party platforms.

  • GPT-6 Sol and Luna

    OpenAI · GPT-6 Model Family

    Released 22 September, carrying GPT-6 Astra's strengths into faster, cheaper models. Sol is the balanced model for interactive and agentic coding needing multistep validation; Luna is the lightweight, lowest-cost model in the family for smaller, faster tasks. Sol drops to $2 and $10 per million input and output tokens from GPT-5.6 Sol's $4 and $20; Luna to $0.10 and $0.50 from $0.20 and $1.20. OpenAI states Sol makes about half as many mistakes as its predecessor on its own factuality evaluation, reaching Astra-level reliability at lower cost. Ships in ChatGPT Work, Codex and the API; Luna also in the desktop app and Free and Go tiers. openai.com returned 403, so this rests on secondary coverage.

  • dots3-note Preview

    Dots Studio (Xiaohongshu) · Dots3 Note Preview

    The first public release in Xiaohongshu's dots3 series: a mixture-of-experts model at 280B total and 16B active parameters, 256 routed plus one shared expert over 46 layers, supporting text, images, video with audio, and audio transcription across a 512K-token context. Released 17 August under Apache 2.0 in BF16 and FP8, deployable via SGLang or vLLM, with a free tier on OpenRouter. Self-reports 78.4% on SWE-bench Verified, 61% on SWE-bench Pro, 79.1% on MMMU Pro and 52.6% on HLE, explicitly without comparing itself to named peer models.

  • Limite 1B - Violetto

    Paradigma · Limite 1B Violetto

    A 1B dense autoregressive transformer trained on under 300B highly curated tokens using synthetic data generation, supervised finetuning and reinforcement learning, built for single-turn mathematical reasoning rather than general assistant use, with context up to 131,000 tokens. Self-reports 74.25% average on BeyondAIME and 94.01% on AIME 2026, with no named-baseline comparison given. Releasing the model, its evaluations, the value model used during training, and a custom vLLM inference plugin, on GitHub and Hugging Face.

  • Claude Opus 5.5

    Anthropic · Claude Opus 5.5

    Released 22 September at $4 and $20 per million input and output tokens, $0.20 for cache read and $5 for cache write, with a faster mode at $8 and $40. Anthropic states it costs 40% less to run than Opus 5 on typical workloads while generating output over 30% faster. Its own benchmark table shows gains over Opus 5 and named competitors Fable 5.1, GPT-6 Astra and GPT-5.6 Sol: 66.4% on Terminal-Bench 4.0 against Opus 5's 52.3%, 67.7% on Humanity's Last Exam with tools against 63.6%, and 81.8% on OSWorld 2.0 against 74.0%. Anthropic also states it attempted to circumvent boundaries around 85% less often than Opus 5. Available on the Claude Platform, AWS, Google Cloud and Azure across Pro, Max, Team and Enterprise. Parameter count and context window are not disclosed.

Vision & Image (5)

  • RADAR

    RADAR-Generalist team, published in Science · Abdominal CT Vision Model

    A generalist vision-language model for abdominal CT diagnosis, trained on more than 400,000 contrast-enhanced CT examinations distilled into 15 million anatomy-aware image-text pairs, learning directly from clinical reports rather than manual annotation. RADAR+ variants finetune further on the Merlin-CT-Train dataset over a UNet vision branch with Chinese and English BERT tokenisers. Released under CC BY-NC-SA 4.0 and explicitly stated as research-only pending clinical validation, with no benchmark scores published on the card itself.

  • Qwen-Image-2.1

    Alibaba Qwen · Open Image Generation

    An open-weight text-to-image and image-editing model with a 7B visual generation component, adding native transparency so a single checkpoint outputs RGBA, multi-reference and local editing, and more realistic textures than the prior version. Ships under the Qwen Research License, non-commercial only, with day-zero support in ComfyUI, Diffusers and vLLM. qwen.ai renders as an empty JS shell to automated fetches, so these details come from secondary coverage rather than the announcement itself.

  • Ming-Image-0.1-Design

    inclusionAI · Design-Focused Image Model

    A 6B text-to-image model specialised for text-rich visual design output such as UI, infographics and posters, with RGBA transparent-background support. Recommended settings are 2048x2048 or 1024x1024 resolution, 12 sampling steps, CFG scale 1.0, in BF16 on a single 80GB GPU. MIT on Hugging Face and ModelScope. The card shows a UI/UX design leaderboard graphic but gives no numeric scores against named competitors.

  • Ovis-Embedding

    ATH-MaaS · Ovis Embedding Models

    A collection of universal omni-modal embedding models: Ovis-Omni-Embedding-3B, Ovis-VL-Embedding-2B and Ovis-VL-Embedding-9B, each tagged for feature extraction. An accompanying paper was published six days before this collection went up. No licence terms or benchmark numbers are stated on the collection page itself.

  • Supra2-IMG

    SupraLabs · Supra2 IMG

    A small text-to-image diffusion transformer at 104.1M parameters, generating 256x256 images using a frozen Flan-T5-Base text encoder, an SD-VAE-FT-MSE, and a 14-layer DiT. Trained for 10 epochs on 5.6M images from the FLUX-Reason-6M dataset on a single H100 in 9 hours. Apache 2.0 with downloadable weights and an inference script. The page claims state-of-the-art quality for its size but gives no benchmark numbers against named baselines.

Video & Animation (2)

  • Calliope

    benjiyaya · Story-to-Video Studio

    A local-first story-to-video pipeline: describe an idea, generate reference images through a local ComfyUI instance, write a screenplay broken into shot clips, then render and stitch the shots with ffmpeg. Everything runs locally over SQLite and folder-based media storage except an OpenAI-compatible LLM endpoint. FastAPI backend, SvelteKit frontend, MIT, 221 stars, actively maintained.

  • WorldCrafter

    ARC Lab (Tencent IEG) & Peking University · Explorable World Model

    Generates explorable, spatially consistent 3D environments from a single image or text prompt, supporting minute-scale navigation and camera-controlled world generation. Rather than explicit depth or geometric warping it uses a camera-queryable implicit memory, a memory encoder plus pose-conditioned readout, to decide which historical frames inform a new view. Evaluated on 145 scenes and 725 camera trajectories, reporting the highest overall VBench score among tested baselines, though the competitor names weren't stated. Code and a fast weights variant released.

Audio & Speech (1)

  • Gemini 3.8 Flash TTS

    Google · Expressive Gemini TTS

    Two new text-to-speech models, Flash and Flash-Lite, improving on Gemini 3.1 Flash TTS with better long-form content, dual-speaker screenplay control and natural-language voice design rather than presets. More than 2,000 ready-made voices, expandable to unlimited custom ones, over 100 languages, voice replication from a 30-second consented sample, and conversational backchanneling. Google cites first place on the Hume AI Voice Design Benchmark at 71.4 and top Voice Arena placements for Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. Rolling out through AI Studio, the Gemini API, Notebook and Vids; no pricing disclosed.

3D & Spatial (2)

  • Mira-Scene

    VAST AI Research · Single-Image 3D Scenes

    Reconstructs an editable 3D scene from a single image through instance segmentation, depth estimation, canonical-coordinate and voxel prediction, object mesh generation, scene assembly and environment-map generation, with resumable multi-stage execution across multiple GPUs. Reports gains over SAM3D using roughly 80,000 training samples and fully automatic annotation against SAM3D's million-plus manually labelled set: Chamfer Distance 0.021 against 0.027, F-Score at 0.1 of 0.843 against 0.817, 3D IoU 0.727 against 0.520. Code and pretrained weights, including a finetuning checkpoint, released.

  • GAE

    HKUST, ARC Lab (Tencent IEG), University of Hong Kong & UT Austin · Geometry-Native World Model

    Embeds geometry directly into the generative latent space, so one shared latent decodes to RGB, depth, camera pose and 3D geometry together instead of treating geometry as an external constraint bolted onto an appearance-only latent. Freezes a geometry foundation model and trains only a compact codec compressing 3,072 channels down to 128, then uses flow matching to generate latents conditioned on text or reference images. Against RAEv2, Stable Diffusion's VAE and the Wan v2.1 video VAE it reports FVD improvements of 12.7% on RealEstate10K and 23.1% on DL3DV, and halves camera trajectory error, ranking first on geometric consistency on 40 of 64 benchmark scenes. Code and an interactive text-to-image demo producing downloadable point clouds are released.

Agents & Automation (4)

  • LibreChat

    LibreChat-AI · Self-Hosted AI Chat

    A self-hosted, open-source ChatGPT-style interface in TypeScript that switches between models from OpenAI, Anthropic, Google, Azure, AWS Bedrock, Mistral, OpenRouter and Vertex AI in one UI. Ships agent building with MCP tool integration, a sandboxed multi-language code interpreter, image generation and editing, web search, speech-to-text, multi-user auth over OAuth2 or LDAP with an admin panel, and inline React and HTML artifact rendering. MIT, 44,982 stars, still actively pushed the same day this was written.

  • Jev Trade

    aowang-ai · Crypto Trading Agent

    A cryptocurrency trading bot for the Hyperliquid derivatives exchange that genuinely uses TypeSafe's Jev, covered here in week 38: it reads Hyperliquid order-book data each tick and queries Jev through TypeSafe's API or Vercel's AI Gateway to decide buy, sell or hold across five isolated BTC, ETH, SOL, DOGE and BNB wallets. Bun and TypeScript backend with a Next.js dashboard, MIT, 155 stars.

  • FLUX 3 Action

    Black Forest Labs · Robot Action Model

    Black Forest Labs' image and video generation line moving into robotics: a 7B world-action model built on the FLUX 3 backbone that jointly predicts future video and robot actions from visual input, trained across gaming, egocentric hand poses, grippers and teleoperation data with flow matching. On the vendor's own RoboLab-120 leaderboard the guidance-distilled variant reports 42.2% success against Cosmos 3 Nano's 36.8% and Pi-0.5's 28.0%, with real Franka arm tests at 93.3% across 30 tasks. Weights released under both non-commercial and commercial licences, with API access.

  • CLM-v0.1-8B

    Contrastive-LM · Agent Decision Model

    Adds two small projection heads, a state head and an action head, on top of a frozen Qwen3-8B encoder, trained with a bidirectional InfoNCE contrastive loss rather than as a generative model, built for scoring, ranking and agent tool or action selection. Trained in three phases: pretraining on roughly 60M Nemotron Q&A pairs, mid-training on 30M synthetic hard negatives, and post-training on 1M agentic trajectories, at a 2,048-token maximum sequence length. Reports state-of-the-art on DeepSWE at 81.6% and Terminal-Bench 2.1 at 87.6%, and claims 4-9x lower latency than Jev (week 38) at comparable zero-shot performance on computer-use and tool-calling. Apache 2.0, weights on Hugging Face, pip installable.

Developer Tools (1)

  • Altar

    Aikido Security · Altar Sovereign Security

    A compressed, open-weight derivative of GLM-5.3, shrunk from 1.51TB to 328GB, a 78.2% reduction retaining 168 of 256 routed experts per layer, via W4A16 quantisation, meant to run entirely inside a customer's own infrastructure, including air-gapped environments, for code review and penetration testing without sending data externally. Deployed on H200 nodes with vLLM, weights published on Hugging Face. Aikido self-reports 60.4% average recall identifying 32 known vulnerabilities across test repositories, aimed at banking, healthcare and industrial operators. No licence name or pricing stated.

Research & Safety (1)

  • Array-Associated Reverse Transcriptases

    Anthropic · AI-Assisted Discovery

    Anthropic reports that Claude agents identified a novel enzyme system, a reverse transcriptase (previously known from jumbo phages) paired with a partner gene and an array of evenly spaced DNA repeats resembling CRISPR arrays, a combination Anthropic says had not been noticed together before. Roughly 950 Claude agents spent 21 hours and 210 million tokens scanning over 200,000 reverse transcriptases across DNA sequence databases, narrowing 3,500 candidate systems to 20, before one agent flagged the tandem repeat pattern. Human lab work has since confirmed the array is expressed as distinct short RNAs, but Anthropic states plainly it does not yet know the function, and this is an early finding shared alongside a preprint rather than a peer-reviewed, function-characterised discovery.

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