Multimodal

OpenAI made its Decisions API generally available for bounded text and image decisions without full responses. It adds native JSON inputs, prompt caching and up to 400 questions per request.

NVIDIA released Nemotron 3.5 Embed and Rerank models with document-image support through Sentence Transformers. NVIDIA reports reranking raises its ViDoRe v3 image-and-text result from 58.82 to 67.56 nDCG@10.

FastVideo releases FastH3 Omni weights and training code, claiming 12× lower latency than base MiniMax-H3. The comparison covers 27 matched workloads on eight GB200 GPUs, and ComfyUI support is available.

FastVideo's FastH3-QAD-v2 and NVFP4 VAE decoders cut weights from 118 to 30 GiB. The team reports generating a five-second 480p clip in 52.7 seconds on one DGX Spark, down from 141 seconds.

Nemotron 3.5 Super VL has 120B total parameters, 12B active per token and a 1M-token context window. Ollama and Factory added the model, with Factory reporting improved tool-call parallelism.

StepFun's multimodal Step 5 Preview is available through OpenRouter and coding tools. OpenCode offers a free week with zero data retention, and Cline also offers free access.

Perplexity's open pplx-embed-v2-late models retrieve text and images in a shared multi-vector space. The 0.6B model can query indexes built by the 9B model, including document pages without OCR.

Google released EmbeddingGemma 2 under Apache 2.0 to map text, code, images, video, and audio into one embedding space. The 740M-parameter version supports 8K context and on-device retrieval.

xAI's experimental TypeScript SDK supports Grok text, voice, image, and video models, plus server-side tools. Code execution runs in isolated, stateless Python sandboxes without network or filesystem access.

FLUX 3 Image supports bounding-box layouts, targeted edits, native 4K output, and ten reference images. Commercial weights are available, open weights are planned, and API use is half-price through October 8.

Perceptron released its 35B Mk1.5 model for drones, quadrupeds, smart glasses, and tool-using agents. The company says the model adds audio, egocentric-video understanding, tracking, and structured outputs, with execution up to 5x faster than Mk1.

CUA released Cua-S1-4B-0.2, a multimodal decision model trained with supervised learning and task-completion RL in live environments. CUA reports 92.9% on a frozen GUI-360 split and released adapters and training code.

Xiaomi released Pro and Flash MiMo-V2.6 mixture-of-experts models with open weights and a 1M-token context window. The release includes an RL dashboard and day-one vLLM support, while training artifacts are planned.

Qwen-Image-2.1 is a 7B model that combines image generation and editing, including native RGBA output and support for up to 10 reference images. It has day-one support in ComfyUI, Diffusers, vLLM-Omni, and Ostris, but its license is non-com

Gemini 3.8 Live and its Extended Thinking variant add bidirectional voice, visual grounding, multilingual speech, and asynchronous tool calls. Google is rolling both models out through its API with audio watermarking via SynthID.

ElevenLabs’ MCP server now lets supported assistants generate speech, music, images, and video. It also exposes transcription, dubbing, sound effects, and tools for refining generated work.

DeepSeek released V4.1 Flash, a 552B-parameter multimodal MoE model with 8B input and 16B output active parameters. It supports up to 1 million tokens of context, while an independent BridgeBench run used 23.5 million tokens on one task.

OpenAI introduced GPT-Image 2.5 Flare and Sunburst API models alongside ChatGPT Images 2.5. The release emphasizes lower generation latency, stronger multi-turn edits, and better reference-subject preservation.

World Labs says Atlas combines visual generation with scene reconstruction. It can reconstruct scenes from images, generate camera-controlled frames, and reframe video.

Meta launched Muse Voice Transcribe, combining streaming transcription, speaker diarization, and endpointing in one model. Meta says it supports code-switching and costs $0.18 per hour.

Google says Gemini Omni 1.1 Flash can reference up to 10 seconds of preceding video when extending a scene, rather than only the final second. It also adds keyframe interpolation and object editing for more controlled video generation.

Perceptron released weights, inference code, and training details for Isaac 0.5, an embodied model for video perception, reasoning, and robot control. The 36B dynamic-MoE model was trained on 1 million hours of video.

WAN 3.0 is now available through Replicate, OpenRouter, and ComfyUI for text-, image-, and reference-driven video generation up to 1080p. Pika says the model supports up to 20 references and 30-second output, while quality and price comparisons remain vendor-reported.

Meta published Muse Spark 1.2 evaluations covering tool-based web page and game creation, robotics planning, and audio-visual tasks. A separate result places it first on Design Arena’s video-to-website benchmark.

OpenRouter is testing Ox Alpha, a free stealth model with a 1M-token context window. The model accepts text, image, and video inputs and is offered with zero data retention during the test.

Sentence Transformers 6.0 adds MultiVectorEncoder for ColBERT-style training, inference, and interpretation, including visual-document retrieval. Index size can increase substantially: 4,874 passages expanded to 608,414 token vectors in one example.

Qwen released its open-weight Qwen3.8 27B vision-language model with 262K native context and adjustable reasoning. In a 484-sample test, enabled-thinking scores fell from above 92% through 64K to 74.3–81.8% at 128K.

LocalLLaMA users said Qwen 3.8 Max 2.4T open weights are text-only while the API keeps vision support. A linked Qwen3.8-27B ModelScope page reportedly returned 404 before release.

Google opened Gemini Omni Flash to developers for video creation and editing from text, image, video, or audio references. The rollout spans Gemini, Flow, AI Studio, the Gemini API, and Vertex AI, with supporting posts showing editing demos rather than independent performance tests.

ComfyUI says Seedance 2.5 is live via Partner Nodes with 30-second runs, up to 50 references, timeline shot control, editing, and multilingual lip sync. Fal, Pika, Venice, and other tools also added access.

MiniMax H3 now has Hugging Face weights, fal endpoints, AI Toolkit LoRA support, and reported single-RTX-5090 local runs. MiniMax also said deployment in the US, EU, UK, and South Korea is available through formal authorization.

MiniMax released H3 weights on Hugging Face for text-to-video, image-to-video, reference-to-video, and editing workflows. vLLM-Omni, ComfyUI, SGLang Diffusion, and fal added support at launch.

Thinking Machines released Inkling-Small, a 276B-parameter MoE with 12B active parameters, multimodal inputs, and a 1M-token context window. Providers added day-zero vLLM, SGLang, Modal, and gateway support.

Google released Gemini Robotics 2, ER 2, and On-Device 2 for humanoid control, embodied reasoning, and on-device adaptation. Demos showed sub-second streaming and multi-robot task handoffs.

MiniMax announced H3 for app and API access, with OpenRouter, Vercel, Pika, fal, Runway, and Venice integrations. Artificial Analysis ranked it No. 1 for video editing, and MiniMax said open weights are planned.

Moonshot published Kimi K3 weights, a technical report, and a blog for a 2.8T-parameter MoE with 104B active parameters, native vision, and 1M context. The license adds separate terms for large model-as-a-service providers.

Alibaba opened Qwen 3.8 Max Preview on Alibaba Cloud, Qwen Chat, Qoder and the web, describing it as a 2.4T model headed for open weights. Early testers praised vision results but disputed coding claims.

Thinking Machines released Inkling with Apache 2.0 weights, 975B parameters, 41B active parameters, text/image/audio support, and up to 1M context. vLLM, SGLang, Modal, Databricks, and Vercel added day-zero support.

Meta said a model scored 30/30 on the APhO theoretical exam. Team posts described data curation, training and live participation, while public posts questioned which model was evaluated.

Goodside and other testers shared chat links where GPT-5.6 Sol, and sometimes Claude Fable 5, hallucinated hidden messages in noise images or meaningless scribbles. Higher effort settings sometimes did better, but failures reproduced.

LingBot 2.0 released code and weights for a real-time world model and robot-action models. The VLA maps 20 robot body configurations into a 55D action format and filters 90,000 raw robot hours to 50,000 training hours.

Riley Goodside tested random-noise and scribble images with no hidden message. GPT-5.6 Sol often produced invented text, while Claude Fable 5 more often refused or identified the image as non-writing.

Meta launched Muse Spark 1.1 in Meta AI and the Meta Model API public preview for coding, tool use, computer use, and multimodal reasoning. Early eval posts ranked it highly while system-card threads flagged safety details.

OpenRouter tested 1,730 visual-reasoning questions across five models and found low-detail images often reduced accuracy while increasing reasoning-token spend. Caps on reasoning effort had the biggest billing impact.

Goodfire introduced Block-Sparse Featurizers, which model activation concepts as multidimensional blocks instead of single SAE directions. The examples cover DINOv3, SDXL, and InceptionV1 activations.

Meta launched Muse Image in its apps and previewed Muse Video from the same media-generation family. Meta says Muse Image can reason, search, write code, self-refine, and use test-time compute before generating images.

X-Humanoid unveiled TG-VLA as a full-size whole-body VLA framework for humanoids, built around HEX, HAF-VLA, and DSRL-DCT. The company claims DSRL-DCT reached 100% success in mobile-manipulation tasks by freezing the VLA and learning a smaller noise-selection policy.

Gemini Omni Flash ranked #1 on Video Arena at 1404 Elo, 101 points above Seedance 2.0 Mini, and ComfyUI posted a text-prompt video-edit workflow. Google noted the leaderboard is third-party, leaving benchmark provenance as the main caveat.

Google shipped Nano Banana 2 Lite for image generation and Gemini Omni Flash for conversational video generation and editing in the Gemini API and AI Studio. The release sets image generation at about 4 seconds and $0.034 per 1K image, while Omni Flash adds multi-turn video edits at $0.10 per second.

Perceptron launched a video_frames input for Mk1 that accepts pre-decoded frames with timestamps instead of forcing clip re-encoding. The change matters for edge and sparse-footage pipelines because 10 minutes of 1080p video can start returning tokens roughly ten times faster.