Quanta AI Labs

35 / Aug 24 – Aug 30, 2026 · 26 RELEASES

AI News Week 35, 2026

Distillation Ate The Week + One Base Model Forked Two Ways. Almost nothing shipped bigger; nearly everything shipped faster. FastH3 cuts MiniMax H3 from 49 denoising steps to four and makes 15 seconds of 768p in under 13 on 8x B200, while fal's H3 Max claims roughly 35x the throughput of MiniMax's own endpoint - the same open base model forked in one week into an open distillation with downloadable weights and a closed API-only product with neither, which is the clearest picture yet of where open weights actually end up. Block3D parallelises text-to-3D into a 5.15x speedup, Fibo-1.5 drops from 50 sampling steps to four, and ByteDance's DiffusionOPSD cuts reward-guided training compute 40% by supervising intermediate denoising steps instead of only the final image. Three open MoE drops landed regardless: GLM-5.3-Flash at 320B with 18B active and the first native multimodality in the GLM-5 line, Qwen3.8-Flash-Next previewing the Qwen4 architecture at 125B with 6B active plus a separate 51B N-gram table, and Tencent's Hy4-preview at 770B with 49B active under Apache 2.0. Google shipped three at once - a speech model at 2.6% WER, a video model that finally extends scenes to 40 seconds, and an LLM-orchestrated geospatial pipeline beating expert baselines on Ebola nowcasting. The quieter thread is provenance and self-hosting: Bria trains Fibo only on licensed images and publishes the compliance claim, while RustFS, SeaweedFS, Twenty and BillionMail each argue from a different angle that the storage, the CRM and the mailing list belong on hardware you own.

Language Models (3)

  • GLM-5.3-Flash

    Z.ai · Multimodal MoE

    The first natively multimodal model in the GLM-5 line: a 320B mixture-of-experts with just 18B active, built on a hybrid of sparse and linear attention. The card evaluates it at context lengths up to 1M — 300K on HLE with tools, 400K on DeepSWE, 164K on BabyVision — and reports Terminal Bench 2.1 at 84.3, ExtractBench Short at 96.3 and Deep SWE at 63.4. MIT-licensed with weights on Hugging Face. Distinct from the GLM-5.3 covered in Week 33: different parameter profile, multimodal from launch, and a different attention design.

  • Qwen3.8-Flash-Next

    Alibaba Qwen · Qwen4 Preview MoE

    Open-weight preview of Qwen's next-generation architecture: a multimodal mixture-of-experts with 125B total and 6B active across 512 experts routed top-10 plus one shared over 48 layers, alongside a separate 51B-parameter N-gram embedding table. It combines Gated DeltaNet with Qwen Sparse Attention, adds a gated residual mechanism and the Muon optimiser, and runs a native 262,144-token context extensible to 1M via YaRN. Reported: GPQA Diamond 91.7, SWE-bench Pro 62.5, DeepSWE 1.1 58.7, AndroidWorld 84.5. Released under qwen-community-1.0; the served Qwen3.8-Flash shares the architecture at $0.16 and $0.47 per million input and output tokens.

  • Hy4-preview

    Tencent · MoE Coding Preview

    Tencent's next Hunyuan generation after Hy3 in Week 28: a 770B mixture-of-experts activating 49B per token across 78 layers — one dense layer then 77 MoE layers of 256 routed experts plus one shared, top-8 routed — with a 10B MTP layer for speculative decoding, Gated DeepSeek Sparse Attention with IndexCache, and a 1M-token context. Tencent's own blind evaluation puts it fractionally ahead of GLM-5.3 at 2.99 against 2.92 and Kimi K3 at 2.99 against 2.94; GPQA Diamond 92.3, SWE-bench Pro 65.7, Deep SWE 64.3. Apache 2.0, with weights and an FP8 variant openly downloadable.

Vision & Image (1)

  • Fibo-1.5

    Bria AI · Fast Text-to-Image

    An 8B-parameter DiT flow-matching text-to-image model, distilled from Bria's FIBO and keeping its structured-JSON caption training and control features. Inference falls from FIBO 1.0's 50 steps at guidance scale 5 to 4-6 steps at guidance scale 1 with no classifier-free guidance, via Distribution Matching Distillation plus a DMD-R refinement stage. It uses SmolLM3-3B as the text encoder and the Wan 2.2 VAE, ships in BF16, and wins a blind pairwise human evaluation over FIBO 1.0 across 247 comparisons. The real differentiator is provenance: trained only on licensed images curated from roughly 1B, with stated GDPR and EU AI Act compliance — though weights are gated under CC BY-NC 4.0 and commercial use needs a separate licence from Bria.

Video & Animation (4)

  • OpenShorts

    mutonby · Video Editing

    Self-hostable video platform bundling three tools: a long-form-to-vertical clipper, a UGC actor generator and a thumbnail and title studio. Whisper transcribes, Gemini picks 3-15 clip moments, and reframing to 9:16 runs through MediaPipe and YOLOv8 face tracking, blurred-background crops, dual-speaker stacks or screencast layouts, with word-level subtitles from faster-whisper and ElevenLabs dubbing across 30+ languages. FastAPI and React with fal.ai for generation; an eight-minute video takes 5-8 minutes on CPU against roughly 50 seconds on GPU. MIT core with 3.7k stars, and a separate commercial licence on the billing directory.

  • Gemini Omni 1.1 Flash

    Google DeepMind · Video Generation

    A production video model update aimed at developers: scene extension in 10-second increments to 40 seconds total using up to 10 seconds of prior context, against a one-second limit previously; first and last keyframe specification to generate continuous in-betweens; a 360p draft mode up to 60% faster at a third of the cost; 1080p and 4K upscaled output; and reference clips of up to three seconds to steer generation. Shipping through AI Studio, the Gemini Enterprise Agent Platform API, Google Flow for AI Plus, Pro and Ultra subscribers, and the Gemini app for scene extension.

  • FastH3 Preview

    Hao AI Lab · Video Generation Speedup

    A distilled, step-reduced student of MiniMax's H3 text-to-video-and-audio transformer, cutting 49 denoising steps to four using Distribution Matching Distillation plus Video Sparse Attention at roughly 90% sparsity in 64-token blocks, with text and audio kept dense. On 8x B200 it generates 15 seconds of 768p in under 13 seconds — 14.38x over base H3 — and 12.88x on a single B200, where latency falls from 678.7s to 47.2s. Trained with 1k+ B200 hours on multi-shot audio-visual data. Preview v1 is text-to-video-and-audio only; weights and a pre-extracted LoRA are on Hugging Face under the MiniMax H3 Community Licence, inference code is open and training code is pending.

  • H3 Max

    fal · Video Generation Speedup

    fal's post-trained, speed-optimised version of MiniMax's open-weight H3, trained and served entirely on NVIDIA GB200 NVL72 systems: five seconds of video in under three seconds, with fal claiming roughly 35x the throughput of MiniMax's own H3 endpoint and citing a third-party Design Arena figure of more than 50x at comparable quality. fal reports it first for overall quality, prompt understanding and aesthetics against twelve rivals including Gemini Omni Flash, Wan 3.0, Seedance 2.5, Kling 3 and Veo 3.1 — its own evaluation. Closed: Playground, Agent and API only, with no weights or code released.

Audio & Speech (2)

  • Gemini 3.5 Transcribe

    Google · Speech Recognition

    Google's speech-to-text model, handling speaker self-corrections inside the transcript, filler-word removal, and multi-speaker identification with timestamps for up to three speakers, plus function calling. Over 85 languages with automatic detection. Google states word error rates of 4.0% streaming and 2.6% non-streaming overall, 5.50% and 5.04% on FLEURS, and time-to-final-transcription 70% faster than its predecessor Chirp 3. In public preview through the Live and Interactions APIs in AI Studio and the Gemini Enterprise Agent Platform, and rolling out to Gboard, the macOS Gemini app and Chrome.

  • VoiceMem

    Tsinghua University · Voice Memory

    A streaming dual-brain memory architecture for voice AI: a left-brain schema and entity store for factual memory, a right-brain model for emotion, preference and personality, fed by streaming ASR, voiceprint, scene detection and emotion recognition. On LoCoMo it reports 91.2% accuracy at K=5 against Mem0 at 61.68%, Zep at 62.93% and EverMemOS at 83.13%, using a mean of 430 memory tokens where Mem0 uses 6,956, at 134ms latency against Mem0's 1,440ms and roughly 300 tokens per turn. Code, a web demo, the ChatMem-400K dataset and weights are all released.

3D & Spatial (5)

  • SplatKit

    mickmumpitz · Gaussian Splatting

    ComfyUI node pack that builds a 3D Gaussian Splatting training set from a single 360-degree equirectangular panorama without leaving ComfyUI. MoGe depth estimation builds a mesh, virtual cameras fly paths through it, a WAN 2.1 i2v model with a Matrix-3D panorama LoRA inpaints the disoccluded regions, then SphereSfM recovers real camera poses and a sparse point cloud, emitting a standard COLMAP dataset any 3DGS trainer can read. Ships 19 nodes and two workflows, a HiRes composite pass that reprojects the source panorama through the reconstructed geometry, and a Triton/PyTorch nvdiffrast replacement needing no compilation. MIT.

  • OVOW

    Tsinghua University & SparcAI · 4D Reconstruction

    One Video, One World turns a single monocular video into a simulation-ready 4D mesh scene with no task-specific training. Four training-free stages chain a vision-language model for instance discovery, category-aware mesh reconstruction for watertight per-instance geometry, iterative render-match-optimise for metric scale and 6-DoF pose recovery, and physics-grounded assembly that enforces contact constraints so the output loads straight into a simulator. Reports 82.7% physical stability under gravity simulation at 3.35 seconds per frame, one to two orders of magnitude faster than the baselines. Code released.

  • FixAnything

    Carnegie Mellon University · Rendering Refinement

    Repairs rendering artefacts in sparse-view 3D representations — 3DGS, NeRF, mesh or point cloud — by fine-tuning Wan2.1-I2V-14B with rank-64 LoRA adapters. Stage one trains jointly across all four representations on paired DL3DV-10K videos using a per-frame preserve/repair mask; stage two applies geometry-aware Flow-DPO preference optimisation with COLMAP-measured camera-pose accuracy as the reward signal. Pose accuracy rises from 61.1 to 68.3 AUC@5 degrees through DPO, and it generalises to MipNeRF-360 and LLFF on roughly 20 paired fine-tuning videos while training about 1% of parameters. ECCV 2026, code released.

  • Block3D

    Zhejiang University · Text-to-3D Generation

    Text-to-3D generation that partitions a discrete 1,024-token shape sequence into contiguous blocks, generates the blocks autoregressively, and denoises tokens within each block in parallel with confidence-guided intra-block correction. Reports a 5.15x end-to-end speedup over a fine-tuned autoregressive baseline — 4.99s against 25.71s — on a held-out split of TRELLIS-500K, at F-score@1% of 0.309 and normal consistency of 0.668. Code is on GitHub; the page states no parameter count, mesh resolution, VRAM requirement or licence.

  • ORBIT++

    Google DeepMind · SfM Benchmark

    An expanded version of the ORBIT benchmark from CVPR 2026 for evaluating structure-from-motion and camera-pose estimation in the wild, built by using 360-degree panoramic video as a source to synthesise challenging perspective clips while keeping ground-truth camera trajectories recoverable. It evaluates MegaSaM, COLMAP and VGGT-Long across six challenge categories — high speed, low texture, low light, crowds, egocentric parallel moving objects and fluids — on ATE and RPE-R, finding low-texture and high-speed scenes hardest overall. Code and dataset released.

Agents & Automation (3)

  • DeepAgents 0.7.11

    LangChain · Agent Framework

    Since Week 27 2025: the project moved from the hwchase17 account to the langchain-ai org and has kept shipping, with 0.7.11 plus companion packages deepagents-code, deepagents-acp and deepagents-talon all bumped in the final days of August. A batteries-included agent harness on LangGraph where every component can be overridden: sub-agent delegation with isolated context windows, local, sandboxed or remote filesystem access, sandboxed shell execution, automatic context summarisation, cross-session memory and human-in-the-loop approval. MIT, 28.7k stars. No major version — this is steady point-release work, not a relaunch.

  • PhoneLLM Alpha 1

    Pipecat · Voice Agent LLM

    Open-weight model built specifically for voice-agent pipelines, based on NVIDIA's Nemotron 3 Nano 30B-A3B — a hybrid Mamba-Transformer MoE at 30B total and 3.5B active, shipped as bf16 safetensors with a 262,144-token context. Pipecat reports single-request P95 time-to-first-token under 100ms on B200. On PhoneBench v1 it claims parity with GPT 5.6 Terra at 94% lower cost and 1,300ms better P95 — though PhoneBench is Pipecat's own LLM-judged benchmark, so the comparison is self-reported on a self-authored test. BSD 2-Clause plus Nemotron attribution terms, with vLLM and SGLang deployment.

  • S1

    Skild AI · Robot Foundation Model

    A robot foundation model built as an in-context learner: it performs manipulation tasks from one or a few video demonstrations with no fine-tuning or post-training, on horizons up to ten minutes — potting a plant, cooking pancakes, pour-over coffee, kit assembly. Pre-trained on 1k-100k hours of teleoperation, egocentric video and simulation. Skild reports 96% accuracy on in-distribution long-horizon tasks and 66% success out of distribution at the 100k-hour scale, a 7x improvement over language prompting alone, and estimates one demonstration is worth roughly 380 post-training examples. Available to commercial partners and early access only.

Developer Tools (5)

  • RustFS 1.0.0-rc.4

    RustFS · Object Storage

    Distributed S3-compatible object storage written in Rust, pitched as a memory-safe Apache-2.0 alternative to MinIO for data-lake and AI workloads. The README claims 2.3x MinIO throughput on 4KB objects under stress on 2-core/4GB nodes, alongside bucket replication and versioning, lifecycle rules, webhook notifications, bitrot protection, Kubernetes Helm charts and OpenStack Swift/Keystone auth beside the S3 API. 31.5k stars; rc.4 landed 28 August and the project is still short of a final 1.0.

  • SeaweedFS

    SeaweedFS · Distributed Storage

    Distributed storage in Go for billions of files, pairing O(1)-lookup blob storage — a central master coordinating volume servers, so metadata is tracked per volume rather than per file — with an S3-compatible API, a POSIX-style Filer mountable over FUSE, and an integrated Iceberg REST catalog for lakehouse work. README benchmarks put a single machine at roughly 15,708 writes/sec on 1KB files and 47,019 random reads/sec on SSD. Apache-2.0, 34.3k stars, with 10.4 erasure coding, rack-aware replication and tiered cloud offload.

  • Twenty

    Twenty · Open-Source CRM

    Self-hostable CRM built on TypeScript and React with a NestJS, PostgreSQL and Redis backend in an Nx monorepo, positioned squarely against Salesforce. CRM objects, views and workflows are exposed as version-controllable configuration, with a CLI for building custom apps on top of the CRM and agent and chat features bundled in. 55.9k stars and 8.8k forks; mostly AGPLv3, with the SDK and UI library under MIT and an enterprise licence on files marked accordingly.

  • BillionMail

    BillionMail · Self-Hosted Email

    Self-hosted email server and marketing platform built on Postfix, Dovecot and Rspamd, bundling newsletter and transactional sending with delivery and engagement analytics plus RoundCube webmail. Installation is an eight-minute script, Docker Compose, or an aaPanel one-click. AGPLv3 with 15,500+ stars, sold on the argument that subscriber data and sending reputation should sit on hardware you control rather than behind a SaaS subscription.

  • Prompt Master 1.8.0

    nidhinjs · Prompt Engineering

    A Claude Skill rather than an application — a SKILL.md plus reference files, with no source code in the repository at all — that rewrites a rough request into an optimised prompt for a chosen target tool. It profiles 30+ systems including Claude, ChatGPT, Cursor, Copilot, Midjourney, Sora and Zapier, extracts nine dimensions of intent, applies one of 13 templates and checks the result against 37 named credit-wasting anti-patterns, with a memory block for cross-session continuity. MIT, and the listed 12.0k stars are unusually high for a three-file repository.

Research & Safety (3)

  • DiffusionOPSD

    ByteDance Seed · Diffusion Training

    On-policy self-distillation for reward-guided diffusion training. Instead of supervising only the final image, it converts image-level reward into bounded positive and negative targets at intermediate denoising steps, refreshing those targets as the policy updates via an exponential moving average — attacking the sparse-supervision problem where endpoint rewards alone had to guide every intermediate prediction. Reports the best score in all 20 reward-matched comparisons across ten metrics against ReFL, FlowGRPO and DiffusionNFT, a 40% training-compute cut on SD3.5-M versus DiffusionNFT, and 63% fewer GPU-hours on a step-distilled Z-Image-Turbo variant. Code released.

  • Planetary Prediction Engine

    Google Research · Planetary Prediction

    An LLM-orchestrated system that runs the entire geospatial prediction workflow from a natural-language query — data discovery, multimodal dataset curation via Population Dynamics Foundation Models and AlphaEarth embeddings, then training gradient-boosted decision trees and MLPs — with no manual data engineering, cutting model building from weeks to minutes. Google reports mean R-squared of 76.8% against 60.0% for expert and Bayesian baselines on US CDC health indicators, 66.1% against 31.5% on Nigerian food-security downscaling, and Recall@10 of 83.3% against roughly 73% on DRC Ebola nowcasting. Described as an experimental capability, with no product or API access.

  • Code World Model

    Westlake University & NTU · World Model

    Splits world simulation in two: a coding agent maintains persistent world state as executable code and emits an intermediate spatiotemporal proxy representation, which then conditions a fine-tuned MiniMax-H3 video model to render the visuals — reasoning about state in code, realising it in pixels. Trained on paired gameplay and real-world video with aligned RGB-proxy annotations from a purpose-built data pipeline. A GitHub link is up, but the project page carries no parameter counts, resolution, latency or benchmark figures, and states no licence.

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