New AI Models

MAPL-EMIT

MAPL-EMIT is an AI model from Google Research and NASA JPL that detects and localizes methane emission plumes globally from NASA EMIT hyperspectral satellite imagery, with findings published in PNAS. Trained on 3.6 million physics-simulated plumes, it cuts through noisy terrain to detect 50% more plumes than human experts and has identified more than 23,000 additional plumes worldwide, including 24 of the 25 largest-emitting landfills. Open-source models are available on Kaggle, inference tools on GitHub, and the global plume database plus visualization app on Earth Engine, supporting researchers, policymakers, and operators in faster, more targeted climate mitigation.

Suno v6

Suno v6 is a new generation of music models developed with industry partners including Warner Music Group, BMG and Believe. For the first time it ships as a suite of three models: the flagship v6 for Pro and Premier subscribers, built for precision and consistently polished results across genres; v6-wild for exploration, producing more varied and unexpected outcomes; and v6-mini, a faster and more efficient model available to everyone. Across the suite, v6 understands more of the musical building blocks musicians use, from vocals and instrumentation to structure and mood, and adds plain-language editing of song sections, multi-source mashups, sampling and beat-building workflows, and creation from text, audio, images and video. Safeguards screen uploaded audio files and lyrics for unauthorized use.

AlphaGenome Atlas

AlphaGenome Atlas is a free genomic research resource from Google DeepMind that precomputes predictions from its existing AlphaGenome model across all roughly 9 billion possible single-nucleotide variants in the human genome. The 1-petabyte catalogue charts molecular effects on gene regulation across hundreds of human and mouse cell types and tissues, and introduces the AlphaGenome Variant Impact (AVI) score, which condenses AlphaGenome and AlphaMissense predictions into a single number for ranking both coding and non-coding variants. Academic researchers can explore it through the alphagenome.google/atlas portal, the AlphaGenome API, and a skill in Google Antigravity.

NV-Reason-CT

NV-Reason-CT is NVIDIA's 3D vision-language model for CT volume analysis, coupling a Qwen3.5-4B language model with a native 3D ViT encoder (Primus 3D ViT) and 3D MRoPE positional encoding, with 4.69B total parameters (4.35B in the active CT pathway) and OpenMDW-1.1 open weights. It takes NIfTI CT volumes auto-cropped to chest or abdomen and resampled to 2 mm isotropic resolution, passing 13,824 visual tokens to the LLM without spatial downsampling, and supports radiology report generation, question answering, and multi-step reasoning with reviewable, disableable step-by-step thinking. The model was trained end-to-end with SFT and GRPO on roughly 550,000 structured QA examples from 70,111 unique CT volumes drawn from CT-RATE, CancerVerse, and an internal NIH dataset, and the model card reports 0.614 Macro-F1 and 0.871 Macro-AUROC on CT-RATE, ahead of VoxelFM, ClinFusion-8B, and MedGemma 1.5. It is intended for radiologists, medical students, and researchers in research and educational settings, not for clinical diagnosis.

ChatGPT Images 2.5

ChatGPT Images 2.5 is OpenAI's new image generation and editing model, producing sharper details, more natural lighting, and richer textures than Images 2.0 while cutting generation latency by up to 50%. It preserves subjects from reference photos more faithfully, follows editing instructions more reliably across multi-turn edits without degrading quality, and handles complex layouts including transparent backgrounds. It is available to ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and web, alongside two API counterparts, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.

GPT-Image-2.5 Flare

GPT-Image-2.5 Flare is one of the two image models OpenAI released in the Images 2.5 API update, bringing the same upgrades in quality, editing, and speed as ChatGPT Images 2.5 to developers. It is the default choice for most applications, delivering higher-quality images than GPT-Image-2 at 50% lower latency, with more natural lighting, richer textures, and better preservation of reference-photo subjects. It follows editing instructions more reliably across multiple turns, so teams can update a single element while keeping the rest of an asset intact. It suits creator and social content, product experiences, visual search, rapid prototyping, and high-volume generation.

GPT-Image-2.5 Sunburst

GPT-Image-2.5 Sunburst is the precision tier of the two image models OpenAI launched alongside ChatGPT Images 2.5, trading longer generation times for an extra level of precision in detailed creative work. Built for premium visual workflows that need tighter control across edits, it suits production-ready campaign creative and polished product imagery, with multi-turn editing, reference images, mask edits, and transparent backgrounds via the Images API and Responses API. Beyond the high tier of earlier GPT Image models, it adds xhigh and max quality settings, both defaulting to auto.

MiMo-X-Flash-Preview

MiMo-X-Flash-Preview is the fast, low-latency preview tier of Xiaomi's new-generation MiMo-X model family, introduced alongside MiMo-X-Pro-Preview through the MiMo Desktop invitation-only beta, where approved users get limited-time free access. While the Pro variant targets heavier reasoning and long-chain execution, the Flash tier prioritizes quicker responses for everyday tasks within the desktop client's smart model routing. The MiMo-X series is optimized for real professional work scenarios, covering complex task reasoning, multi-Agent collaboration, large-scale programming projects, Office documents, web and front-end design, video editing, 3D generation, music generation, and computer control. It remains in preview testing, and its performance may change with future iterations.

MiMo-X-Pro-Preview

MiMo-X-Pro-Preview is the Pro tier of Xiaomi's new MiMo-X model series, unveiled alongside MiMo-X-Flash-Preview as a preview release. At launch it is offered through the Xiaomi MiMo desktop client under a limited-time, invitation-based free program, and Xiaomi has not yet disclosed parameter counts, architecture details, or benchmark results. The model is positioned for all-round task delivery, targeting complex task reasoning and real-world work scenarios such as office work, design, coding, and omni-modal creation.

MAI-Image-2.6-Flash

MAI-Image-2.6-Flash is the speed-optimized variant of Microsoft's MAI-Image 2.6 family, built to deliver quality comparable to the base MAI-Image-2.6 model for latency-sensitive, high-throughput production workloads. Microsoft reports it generates images 2.8x faster than GPT-Image-2-Medium with 72% greater efficiency, with a median response time of 12.2s at 1024x1024 under a 100 RPM load test, and cites an Image Edit Arena Elo of 1311 at $19.50 per 1,000 images on the quality-cost frontier. It shares the family's multi-image reference editing, web grounding, dynamic aspect ratios and up to 1.5K resolution support, and is available to developers in Microsoft Foundry.

Nemotron-Labs-3-Competitive-Coding (550B A55B)

NVIDIA Nemotron-Labs-3-Competitive-Coding is a competitive-programming specialist LLM fine-tuned from Nemotron-3-Ultra-550B-A55B, built on a Mamba2-Transformer hybrid LatentMoE architecture with Multi-Token Prediction, 550B total and 55B active parameters, and a 262,144-token context. It was trained with a single-epoch SFT run on 477,642 synthetic reasoning traces distilled from GLM-5.2 across 22,000 curated contest problems, with open weights under the openmdw-1.1 license released in NVFP4 post-training quantization. With the iterative GenCorrect test-time strategy it scored 502.0/600 on IOI 2025, solved 9.6/12 problems on ICPC 2025, reached 74.5% Pass@1 on LiveCodeBench Pro, and a live run under official IOI 2026 rules scored 535.4/600, above the top human contestant's 498.27. It targets competitive programming, algorithmic problem solving, and code-reasoning research, and is not intended for general-purpose chat or production coding assistants.

WeatherNext 3

WeatherNext 3 is the third-generation global weather forecasting foundation model from Google DeepMind and Google Research, ranked the most advanced and accurate global weather model to date in Brightband's independent live evaluations. Instead of relying on numerical weather prediction output, it learns directly from real-time observations, using raw geostationary satellite data to generate high-resolution forecasts every hour, about five times sharper than WeatherNext 2 with surface variables down to 5 km. It also delivers markedly more precise precipitation forecasts (up to 60% CRPS improvement against IMERG) and new clean energy variables such as 100-meter wind speed, cloud cover, and solar radiation, and now powers weather features across Google Search, Gemini, Maps, and Google Cloud.

MAI-Transcribe-2

MAI-Transcribe-2 is Microsoft AI's speech recognition model for turning noisy, real-world audio into precise transcripts across 60 languages. It averages 5.2% WER on the FLEURS benchmark to rank first, and Microsoft says it defines the accuracy and latency Pareto frontier on Artificial Analysis while processing long-form audio up to 10x faster than leading competitors. The model adds speaker diarization, word-level timestamps, keyword biasing, verbatim or clean transcription styles, code switching and automatic language identification, at $0.10 per hour of audio. It targets clinical note-taking, legal documentation, captioning and accessibility workloads.

Nemotron-3-Labs-Ultra-Math

Nemotron-3-Labs-Ultra-Math is NVIDIA's open mathematical reasoning specialization of Nemotron 3 Ultra, offered as SFT and RL checkpoints (called Nemotron-3-Ultra-SFT and Nemotron-3-Ultra-RL in the technical report An Open Recipe for IMO Gold). Built on the 550B total, 55B active hybrid Mamba2-Transformer LatentMoE model with Multi-Token Prediction and up to a 1M token context, it is trained to solve hard mathematical problems and identify mistakes in proofs, and was deployed as part of an ensemble system that achieved a gold-medal-level score at the International Mathematical Olympiad 2026. The SFT stage learns from about 39B tokens of proof generation, refinement and verification traces, and the RL stage trains on 9,597 proof-generation prompts from hard AoPS problems. Released under the OpenMDW-1.1 license for commercial and non-commercial use, it targets researchers and developers working on mathematical reasoning and proof verification.

GPT-6 Astra

GPT-6 Astra is an OpenAI frontier model, now generally available on the API in a limited rollout after its early Astra announcement. OpenAI positions it as its strongest model for coding, computer use, research, and long-horizon multi-step tasks, scoring 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench. It is the first model to meet the Critical cybersecurity capability threshold under the Preparedness Framework, able to find previously unknown vulnerabilities and build exploit chains in hardened systems without step-by-step human guidance, with early alpha access expanding to defensive users through Daybreak Blue under stronger safeguards and misalignment monitoring.

GWM Worlds 2

GWM Worlds 2 is Runway's real-time interactive world model, available as a Research Preview. Built as an autoregressive diffusion video and audio model, it continuously generates 720p video at 24 fps with 48,000 Hz audio, and sessions have no preset length. Users define a world via the WorldPrompt format (a genesis prompt, first frame and timestamped event stream), then steer it with free-form text actions addressed to any subject or the scene itself, alongside continuous camera motion. Runway points to uses in interactive entertainment, filmmaking, visual novels, agentic simulation and generative interfaces.

Gemini 3.8 Flash

Gemini 3.8 Flash is Google's Flash-tier workhorse LLM, released September 2, 2026 as the company's best reasoning and coding model yet at the speed and price of 3.7 Flash. It scores 54.9% on HLE-Verified and 47.2% pass@1 on CWE-Bench, nearly matching a leading frontier model at a fraction of the cost, with introductory pricing of $0.75 per million input tokens. It targets long-horizon coding, autonomous agents, and enterprise workflows; a sibling variant, 3.8 Flash Cyber, serves vetted security defenders.

Gemini 3.8 Flash Cyber

Gemini 3.8 Flash Cyber is Google's most capable cybersecurity model, released September 2, 2026 alongside Gemini 3.8 Flash on the same foundational intelligence, with defensive capability prioritized. It targets autonomous vulnerability discovery and automated patching: it surpasses 3.5 Flash Cyber and much larger frontier models on CyberGym, scores 47.2% pass@1 on CWE-Bench versus 47.8% for a leading frontier model at higher cost, and produced 2.6x more correct Chrome vulnerability patches than larger commercial models. Access is application-only via the Fairwind Program for trusted defenders.

Muse Spark 1.3

Muse Spark 1.3 is Meta's newest LLM, succeeding Muse Spark 1.2 with stronger agentic and coding performance. Meta's benchmark scorecard pits it against Muse Spark 1.2, GPT 5.6 Sol (max), and Opus 5 (max) across agent, coding, instruction-following, and long-context evaluations; in internal engineer comparisons it uses ~20% fewer tool calls and ~25% fewer tokens than 1.2. It also handles long-horizon instruction following and multitasking better and resists prompt injection. Available now in Muse Code and Meta Model API; a max reasoning mode and open weights release are planned.

Qwen3.8-Max-0902

Qwen3.8-Max-0902 (alias qwen3.8-max-2026-09-02) is the dated snapshot of Qwen3.8-Max, frozen at that time point with capabilities matching the base model. It keeps the 1M token context window, thinking mode and the full tool ecosystem, and accepts text, image and video input while outputting text. The official page highlights engineering-grade coding for complex projects and long-range autonomous development, stronger collaborative agentic tool scheduling, and refined vision understanding, suiting it for hard coding tasks, multi-tool agent workflows and document or chart comprehension.