Search “GitHub”

Google DeepMind Releases Genie 4 Real-Time World Model with Native Audio Generation

Google DeepMind has released the Genie 4 world model as a research preview to trusted testers; it can generate explorable worlds in real time from text and image prompts, and natively generate ambient sound, music, and character dialogue. The model introduces Gemini 3.8 Flash as a planner within the simulation loop, responsible for breaking down goals, inspecting frames, and replanning, which DeepMind calls an early step toward hybrid world models. Genie 4 is hosted only on Google's platform and will later be opened to AI Ultra users and Gemini API partners; all video and audio carry SynthID watermarks.

OpenBMB Open-Sources MiniCPM-OCR Suite, Featuring a 1B Parsing Model and a 2B Verification Model

OpenBMB has released the open-source MiniCPM-OCR suite, comprising the 1B MiniCPM-OCR 1.0 and the 2B decision model MiniCPM-OCR Jev. MiniCPM-OCR 1.0 scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2; Jev can perform cross-model verification of text, formulas, and tables, scoring 93.6 on JevOCRBench, surpassing JEV-35B-A3B's 90.1 and approaching Qwen3.5-397B-A17B's 93.8. Powered by DSpark, its speed is 1.49 times that of PaddleOCR-VL-1.6 and 3.62 times that of MinerU2.5 Pro; with SDK concurrency optimization, it can reach 128 pages per minute, and it supports deployment on vLLM, SGLang, llama.cpp, and Ollama.

Google DeepMind releases Gemini 4 Carbon, focused on cybersecurity defense and coding

Google DeepMind has launched Gemini 4 Carbon, calling it a frontier model for long-horizon agentic coding and cybersecurity defense, ranking first with 59 points on the Artificial Analysis Intelligence Index v5.0. The model achieved 96.2% on the GraphWalks long-context test, and 88.2%, 54.8%, and 71.0% on SEC-bench Pro, ExploitGym, and CWE-bench, respectively, but on OSWorld-2.1 it trailed Claude Opus 5.5's 87.2% with 81.5%. Carbon is currently available only to vetted cybersecurity defenders, with introductory pricing of $2 per million input tokens and $10 per million output tokens, later $4 and $20; the timeline for opening it to developers and the public has not been announced.

NVIDIA Releases Nemotron 3.5 Transcribe: An Open-Source 30B MoE Speech Model…

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight large speech model that supports 40 languages and up to 1 hour of audio per pass, and can select output formats such as verbatim, written, speaker-labeled, contextual, and translation via text instructions. The model has approximately 34B total parameters and about 4.3B activated parameters, adopts a hybrid Mamba-Transformer MoE architecture, can run in BF16 on a single H100 80GB, and the NVFP4 version at about 21GB can run on an RTX 5090. NVIDIA self-reports an average WER of 4.71% across 9 public English Open ASR test sets, the best result among open-weight models, though commercial APIs still lead; its shortcomings are that it only supports offline operation, and the Chinese far-field meeting cpCER is 34.95%.

Block Releases Dessa, a Fully Duplex Speech Model Trained Purely on Speech

Block's AI research lab has released Dessa, a full-duplex speech model, opening a waitlist in the form of a research preview. It is trained only on human voice conversation audio, without using text or pretrained models. The model can run locally on a MacBook Pro M4 Max or higher configuration, faster than real time; training used one NVIDIA GB200 NVL72 (72 GPUs). No benchmark scores, parameter count, weights, paper, or API have been published so far.

Baseten Releases 2B Decision Model joey-1, Priced at $0.01 per Million Tokens

Baseten's Base Labs has released a 2B decision model, joey-1, post-trained on Qwen3.5-2B, which directly returns labels and calibrated probabilities instead of generating text, targeting high-concurrency routing, classification, and guardrail checks. The model scores 63.30 on the public Jev Decision Index 0.4 set, the highest among models under 4B, but 8.94 points lower than the 9B elephant-1. Model APIs are priced at $0.01 per million input tokens, with a P50 latency of 20 ms for an average 2000-token request; the weights are open-sourced under the MIT license on Hugging Face and can be self-hosted on a single CUDA GPU.

OpenAI Releases GPT-6.1 Terra, API Price at Half of GPT-6.1 Sol

OpenAI has released GPT-6.1 Terra, replacing GPT-5.6 Terra, with API pricing at $1 per million input tokens and $5 for output, half that of GPT-6.1 Sol. The model supports a 1.05M token context and a maximum output of 128K, targets visual reasoning, computer operation and writing, and is offered only through OpenAI-hosted API, ChatGPT Work and Codex. On Agents' Last Exam v1 it scores 60.0% at the highest effort tier at $1.01 per task, and 85.5% on MMMU-Pro vision, but only 50.3% on Terminal-Bench 4.0, trailing GPT-6.1 Sol and Claude Sonnet 5.5.

Moonshot AI releases Kimi K3.1 Pro: 2.8T total parameters, 104B activated parameters, weights scheduled to be opened on November 4

Moonshot AI releases Kimi K3.1 Pro, a multimodal MoE model with 2.8T total parameters, 104B activated parameters, and a 1 million token context window. The Standard and Highspeed versions are now available on the Kimi website, App, Kimi Work, Kimi Code and API. The company says Highspeed offers 2–3x faster output at twice the price of the Standard version. Artificial Analysis ranks it fourth on the Intelligence Index with a score of 55, behind Opus 5.5's 58; on Arena's Code Arena WebDev it ranks 8th with 1692 points, up nine places from K3. Standard pricing is $2 per million tokens for input, $10 for output, and $0.2 for cached input, one-third lower than K3; the weights have not yet been released, and the company promises to open them on November 4.

Hugging Face Releases Open-Source DNA Model Carbon-X, Runs Locally on a Laptop

Hugging Face on Thursday released the open-source DNA understanding model Carbon-X, which can run locally on a laptop, letting users ask in natural language how DNA changes affect gene activity and health. The model was developed in collaboration with researchers at Zhongguancun Academy and the University of Naples Federico II; it extends the vocabulary of an existing language model to read DNA in six-letter units and was further trained on 300 billion tokens of DNA sequences, English text, and samples pairing DNA with biological descriptions. The model has 14B parameters and can process sequences of about 50,000 base pairs at a time; it cannot upload a complete human genome. The developers also note that the prediction accuracy has not yet been validated through wet-lab experiments.

English translation is being prepared.

NVIDIA Releases Nemotron 3.5 Series of Open Models, Covering Embedding, Reranking, and Document Parsing

On October 21, 2026, NVIDIA expanded the Nemotron Retriever series, launching Nemotron 3.5 Embed (8B BF16, 1B BF16, 1B NVFP4), Nemotron 3.5 Rerank 8B BF16, and the 631-million-parameter Nemotron 3.5 Parse, aimed at multimodal retrieval and document processing for enterprise agent workflows. NVIDIA says that adding Rerank to the agent retrieval pipeline reduces the token cost per query by about 14% (with the 8B embedding model) and 13% (with the 1B model), respectively, while also improving retrieval accuracy. Parse scores 75.20 on ParseBench and processes 10.5 pages per second on a single H100; the open weights, data, and training recipes have been released on Hugging Face and can be deployed via NVIDIA NIM as well as Baseten, DeepInfra, and FriendliAI.

English translation is being prepared.

English translation is being prepared.

Reflection open-sources 501B-parameter MoE model Beam under Apache 2.0

On October 16, Reflection released its first open-weight model Beam on Hugging Face under the Apache 2.0 license, and Baseten provided day-0 support the same day through Model APIs and Dedicated Inference. Beam is a 501B total-parameter MoE autoregressive Transformer with 23B activated per token, a context window of up to 1 million tokens, pretraining on 23.8 trillion tokens, and weights released in BF16, FP8, and NVFP4 precisions. The evaluation results given by Reflection are GPQA Diamond 90.5%, SWE-bench Verified 80.9%, AIME 2026 97.8%, Terminal-Bench 2.1 80.1%, and MCP Atlas 78.7%, and it also open-sourced the terminal agent tool Mirror CLI.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.