Quanta AI Labs

22 / May 25 – May 31, 2026 · 28 RELEASES

AI News Week 22, 2026

Opus 4.8 + Local Frontier — Anthropic ships Claude Opus 4.8 (1M context, 88.6% SWE-bench Verified, 74.6% Terminal-Bench, best computer-use model). StepFun releases Step 3.7 Flash (198B MoE, ~97% of Opus 4.6 quality at 1/9 cost). OpenBMB ships MiniCPM5-1B (Think/No-Think dual mode, 131k context, claimed 1B-class SOTA). PrismML ships Bonsai Image 4B in 1-bit and ternary variants — 6s on M4 Pro, 9.4s on iPhone 17 Pro Max. NVIDIA NVlabs LongLive 2.0 brings NVFP4 training + inference for long video (45.7 FPS on 5B). DeepSWE benchmark exposes 70% solve-rate ceiling for current frontier coding agents. Big 3D wave: TriSplat (feed-forward triangle splatting), GenRecon (scene-level Trellis.2), PhysX-Omni (unified rigid/deformable/articulated), CubePart (Roblox SIGGRAPH 2026 part-controllable), Pantheon360 (CVPR 2026 360° digital twins). Video research: Soap2Soap (multi-agent cinematic remaking), SCOPE (FPS world models), InstructAV2AV (joint audio-video editing), ControlLight (continuous illumination control). Agent research: AutoScientists (decentralized multi-agent biomedical research, +8.33% on BioML-Bench), BES bidirectional evolutionary search, γ-World multi-agent world modeling. Relighting: Relightable Holoported Characters (MPI), PIXLRelight (intrinsic-conditioned). Tooling: AionUi (20+ CLI cowork app), Harness open-source DevOps platform, claude-code-harness (autonomous Plan→Work→Review), Surya OCR (90+ languages), guy-hartstein company-research-agent, datacurve DeepSWE. Vision-Language: NVIDIA LocateAnything (parallel box decoding, 138M samples). Pixel diffusion: NVIDIA PiD (210ms 2K decoding on GB200).

Language Models (4)

  • rahul238xaviers/large-language-model

    rahul238xaviers · Code Generation

    Custom 1.6B-parameter decoder-only GPT trained from scratch for Rust code completion on Apple Silicon using MLX. Includes a Gradio playground and a multi-chapter training journal documenting OOM fixes on M3 Ultra and GQA acceleration.

  • Claude Opus 4.8

    Anthropic · Reasoning

    New Anthropic frontier model with 1M-token default context on API/Bedrock/Vertex, 128k max output and adaptive thinking. Scores 88.6% SWE-bench Verified, 69.2% SWE-bench Pro, 74.6% Terminal-Bench 2.1 and 84% Online-Mind2Web (best computer-use model tested); Fast mode is ~2.5× faster and 3× cheaper than Opus 4.7.

  • Step 3.7 Flash

    StepFun · MoE Architecture

    198B-parameter MoE vision-language model pairing a 196B language backbone with a 1.8B vision encoder — activates ~11B params/token, 256k context, three reasoning levels, ~400 tok/s throughput. With Advisor Mode hits 97% of Claude Opus 4.6's SWE-bench Verified score at ~1/9 the per-task cost ($0.19 vs $1.76).

  • MiniCPM5-1B

    OpenBMB · Small Language Models

    Dense 1.08B-parameter LlamaForCausalLM with 24 layers, GQA (16 query / 2 KV heads) and 131,072-token context, built for on-device deployment. Single checkpoint supports both Think and No-Think modes via `enable_thinking`. Claimed 1B-class open-source SOTA on agentic tool use, code and reasoning. Apache 2.0.

Vision & Image (7)

  • LocateAnything

    NVIDIA · Visual Understanding

    Unified vision-language grounding and detection model that replaces token-by-token coordinate generation with Parallel Box Decoding — emits full boxes/points as atomic units in a single decoding step. Trained on the 138M-sample LocateAnything-Data covering grounding, GUI/doc understanding, dense detection and OCR localization.

  • ControlLight

    arXiv · Image Restoration

    Controllable low-light enhancement framework with continuous illumination-strength supervision. Introduces a misalignment-aware weighted flow-matching loss to preserve structure across enhancement strengths and ships a large-scale real-world degraded-image dataset.

  • PiD (Pixel Diffusion Decoder)

    NVIDIA Spatial Intelligence Lab · Text-to-Image

    Pixel-space diffusion decoder replacing the VAE in latent diffusion pipelines — converts 512² latents to 2048² in <1s on RTX 5090 (210ms on GB200), ~6× faster than cascaded SR. DMD2-distilled to 4 sampling steps; works with VAE and semantic latents (SigLIP, DINOv2).

  • Bonsai Image 4B

    PrismML · Text-to-Image

    1-bit and ternary compressed variants of a 4B-class diffusion transformer (modified FLUX.2 Klein 4B) — memory drops from 7.75GB to 0.93GB (1-bit) or 1.21GB (ternary), up to 8.3× reduction. Retains 88% / 95% of FLUX.2 Klein 4B quality; generates 512² in ~9.4s on iPhone 17 Pro Max, ~6s on M4 Pro. Apache 2.0.

  • SEGA

    arXiv · Text-to-Image

    Training-free resolution-extrapolation method for Diffusion Transformers (Spectral-Energy Guided Attention) — makes RoPE component scaling frequency-aware, adapting it to the spectral structure of the current latent. No retraining or architecture changes; improves structure/semantics/visual quality at higher resolutions on Flux and Qwen-Image.

  • PIXLRelight

    arXiv · Image Editing

    Feed-forward single-image relighting bridging PBR and learned image synthesis via shared intrinsic conditioning (albedo, diffuse shading, non-diffuse residuals) — obtained from either real photographs at training or path-traced PBR renders at inference. Transformer renderer applies target illumination through per-pixel affine modulation.

  • surya

    datalab-to · OCR & Document

    Open-source document AI toolkit covering OCR, layout analysis, reading-order detection and table recognition across 90+ languages. Used as the OCR backbone for Datalab's Marker PDF-to-Markdown pipeline.

Video & Animation (6)

  • LongLive 2.0

    NVIDIA NVlabs · Text-to-Video

    First NVFP4 parallel training + inference system for long video generation. Balanced sequence-parallel autoregressive training, W4A4 NVFP4 inference on Blackwell GPUs, async streaming VAE decoder — 2.15× training and 1.84× inference speedup; 5B variant hits 45.7 FPS.

  • Soap2Soap

    Show Lab (NUS) · Video Editing

    Multi-agent framework for series-level cinematic remaking (full-episode stylization or actor replacement across hundreds of shots) using scene-aware JSON screenplay, dynamic visual reference anchors, and batch keyframe generation in a grid-based shared latent context. Tackles identity drift and background mutation in long video-to-video.

  • InstructAV2AV

    arXiv · Video Editing

    First end-to-end framework for open-world joint audio-video editing from text instructions, built on a foundational AV generation model with source-instruction gated attention. Trained on InsAVE-80K source→target AV pairs; preserves AV synchronization where prior video-editing pipelines ignored audio.

  • SCOPE

    arXiv · Video Understanding

    Interactive FPS-game world model built on a pretrained video diffusion with per-pixel temporal action conditioning, handling a hybrid action space (4D dual-joystick + 6 binary buttons) every frame. Trained on 7 FPS titles via the CrossFPS dataset; generalizes zero-shot to unseen games.

  • γ-World (Gamma-World)

    NVIDIA + Tsinghua · Real-time Video

    Generative multi-agent world model running at 24 FPS real-time via Simplex Rotary Agent Encoding (parameter-free 3D-RoPE extension giving each agent a distinct phase) and Sparse Hub Attention reducing cross-agent attention from quadratic to linear. Trained on 2-player data, generalizes zero-shot to 4 independently controlled players.

  • Pantheon360

    Bosch US (CVPR 2026) · Text-to-Video

    Controllable 360° video diffusion for digital-twin generation from sparse 360° inputs. Uses an explicit 3D Cache reconstructed from the input as a geometric scaffold for any user-defined camera path, so the diffusion model only handles texture refinement while the cache enforces global geometric consistency.

3D & Spatial (5)

  • TriSplat

    ZIP Lab (Zhejiang University) · 3D Reconstruction

    Feed-forward 3D reconstruction network outputting oriented triangle primitives directly from sparse unposed images in a single forward pass — predicts point maps, triangle attributes, camera poses and intrinsics jointly. Output meshes load directly into physics engines and renderers without post-hoc mesh extraction.

  • GenRecon

    TUM (Niessner Lab) · 3D Reconstruction

    Multi-view 3D scene reconstruction cast as conditional 3D generation over overlapping spatial chunks, lifting Trellis.2's object-level prior to scene scale. Outputs editable PBR meshes for relighting/editing via a projection-based conditioning that's permutation-invariant to view order.

  • PhysX-Omni

    S-Lab NTU + ACE Robotics · 3D Generation

    Unified simulation-ready physical 3D generation framework for rigid, deformable and articulated objects, using a VLM-tailored geometry representation that encodes high-resolution structure without compression. Releases PhysXVerse dataset and PhysX-Bench covering geometry, scale, material, affordance, kinematics and description.

  • CubePart

    Roblox (SIGGRAPH 2026) · 3D Generation

    Two-stage diffusion 3D generator built on a VecSet latent shape representation — takes a global text prompt + open-ended list of part names and emits one mesh per part that assembles into a coherent object. Separates global shape synthesis from part-level decoding; outputs drop straight into game engines.

  • Relightable Holoported Characters

    MPI Informatics (GVV) · 3D Reconstruction

    Person-specific free-view rendering and relighting of dynamic full-body humans from sparse RGB videos. Transformer-based RelightNet predicts relit appearance in a single pass (no OLAT basis), trained on a new lightstage capture that interleaves random-environment-map frames with uniformly lit tracking frames.

Agents & Automation (3)

  • AionUi

    iOfficeAI · Coding Agents

    Free open-source local desktop app wrapping 20+ AI coding CLIs (Claude Code, Codex, OpenClaw, Hermes Agent, Gemini CLI, Qwen Code, Goose) into a single Cowork workspace with auto-detection. Stores all session data in local SQLite — no server upload.

  • company-research-agent

    guy-hartstein · Agent Orchestration

    Agentic company due-diligence tool built on LangGraph + Tavily, using Gemini 2.5 Flash and GPT-5.1 for multi-agent inference over public web data. Produces structured diligence reports from a company name input.

  • AutoScientists

    Harvard (Marinka Zitnik Lab) · Agent Orchestration

    Decentralized (no-orchestrator) multi-agent system for long-running computational science that alternates discussion (forming teams around hypotheses) with parallel execution. Hits 74.4% mean leaderboard percentile on BioML-Bench (24 tasks across imaging, protein engineering, single-cell omics, drug discovery) — +8.33% over the strongest prior biomedical agent.

Developer Tools (2)

  • Harness Open Source

    Harness · AI Infrastructure

    End-to-end open-source developer platform bundling source-control management, CI/CD pipelines, hosted dev environments and artifact registries into one self-hostable system. Positioned as open alternative to GitHub + Jenkins + JFrog stacks.

  • DeepSWE

    Datacurve · Evaluation & Benchmarks

    Coding-agent benchmark of 113 hand-authored long-horizon tasks across 91 repos and 5 languages (TS, Go, Python, JS, Rust) using Harbor format with Docker envs and behavioral verifiers (0.3% FP / 1.1% FN). Hand-written from scratch to avoid pretraining contamination — current leader gpt-5.5[xhigh] hits 70% solve rate.

Research & Safety (1)

  • Bidirectional Evolutionary Search (BES)

    Harvard (Yilun Du et al.) · Research Paper

    Self-improvement search framework coupling forward candidate evolution (recombining partial trajectories) with backward goal decomposition into checkable subgoals for dense feedback. Improves on best-of-N and tree search where mainstream post-training fails; tops open-source frameworks on three open problem-solving benchmarks.

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