Odyssey-3 predicts how a generated environment changes as users move the camera, take actions or trigger events. The preview supports first-person and third-person navigation as well as independent camera movement. According to Odyssey’s announcement, the Pro version scored 66.1 on the Physics-IQ Verified video-to-video benchmark, a record for that leaderboard.
同一事件，精选展示《Odyssey 发布世界模型 Odyssey-3，Flash 版本已开放免费体验》Ultrafast is rolling out for GPT-6.1 Sol in the API, Codex, and ChatGPT Work. The stated speed is up to 8 times that of Sol Standard, while the cost is 6 times higher; the maximum speed is not a guarantee for every use.
同一事件，精选展示《GPT-6.1 Sol Ultrafast 开始推送，模型指令调整响应也获改进》GPT-6.1 Sol Ultrafast is rolling out today in the API, Codex, and ChatGPT Work. The announcement claims speeds up to 8 times faster than Sol Standard with near-Astra intelligence, and says the model now responds more quickly to real-time changes in direction.
Odyssey-3 predicts subsequent frames from earlier ones and incoming actions, allowing generated scenes to respond to movement. Its Flash version turns a text prompt into an explorable world that the source says anyone can try for free today. The Pro version has the highest reported Physics-IQ Verified video-to-video score, 66.1.
Odyssey-3 predicts subsequent frames from earlier frames and incoming actions, allowing scenes to respond to movement. Odyssey says its distilled Odyssey-3 Flash is available to try for free; the Pro version's reported score on the Physics-IQ Verified video-to-video benchmark is 66.1, with no independent verification provided in the material.
同一事件，精选展示《Odyssey 发布世界模型 Odyssey-3，Flash 版本已开放免费体验》JetBrains has released Mellum2.1, a 12B-parameter mixture-of-experts model that activates 2.5B parameters per token, with weights released under Apache 2.0. In JetBrains’ self-reported evaluation using a shared pipeline, its SWE-bench Verified score rose from its predecessor’s 2.0 to 47.0, but remained below Qwen3.5-9B’s 50.0.
Preferred Networks has released PLaMo 3 Translate 31B, expanding language support from Japanese and English in the previous version to 53 languages and adding an online meeting translation mode covering 15 languages. PFN says its average score across four translation benchmarks exceeds those of GPT-6.1 Sol and GPT-6 Astra; input costs 14 yen per 100,000 characters.
The 0.6B and 9B models share an embedding space for retrieving text, images and rendered PDF pages. Their weights are available on Hugging Face under the MIT license for self-hosting; a Perplexity-hosted API is planned but not live. Perplexity reports a 92.4% MADQA score for the 9B model, but all scores are self-reported and the technical report is not yet available.
OpenAI says GPT-6 Luna is rolling out to ChatGPT Free and Go users from October 8, while paid users are receiving GPT-6 Sol. The updated chat interface can generate interactive diagrams and tools for a question. OpenAI’s deployment safety report also shows Luna regressed from its predecessor on some safety evaluations involving minors.
OpenAI announced the GPT-6 series and an intelligent UI for all ChatGPT users. The interface can combine text, images and interactive components in its responses; one example uses interactive visuals to teach mahjong. GPT-6 Astra, Sol and Luna were previously available only through ChatGPT Work and Codex.
Liquid AI has released d1-3B and d1-omni-600M, which accept text and images, and text with either an image or audio, respectively. They return probabilities for predefined answers in one pass without generating text, for tasks such as request classification and moderation. Both checkpoints are available on Hugging Face, and the license allows free commercial use below $10 million in annual revenue.
Falcon-ASR focuses on the Emirati dialect and also supports English, French, Spanish and Portuguese. TII says it achieved an average word error rate of 20.92% across six Arabic test sets, versus the best published result of 23.17% in the leaderboard snapshot it used. A Hugging Face demo is available; API access and native applications are planned.
@smallest_AI's Pulse ranks first among seven systems in Voice Arena's diarization and ASR track, with a 24.4% diarization error rate versus 40.7% for runner-up ElevenLabs Scribe v2. The benchmark uses far-field recordings of real conversations as input and labels built from a separate microphone on each speaker.
Google has released Nano Banana 2.1, an image generation and editing model, and says it improves visual design, mask-based editing and subject consistency. Users can select and change part of an image without regenerating the whole picture. The model is rolling out gradually to the Gemini app, AI Mode in Google Search and other products.
Musubi released PolicyLM-1.7B, a lightweight decision model for real-time content moderation, with open weights. The company says it can apply policies written in plain English to messages in under 50 milliseconds and does not need retraining when a policy changes.
Mistral AI CEO Arthur Mensch said the company would release a new AI model today and claimed it outperforms Chinese models in certain areas, including cybersecurity. He did not specify which models were compared or identify the other areas.
Reka has released Rho-1, a 19B model trained from scratch. It can process text, images and video and output robot actions; Reka showed it drawing an image, animating and editing it, then explaining the result across five turns without tools or another model. The research preview is available by email request; public weights, an API and pricing are not yet available.
Reflection AI has released Beam, its first open-weight model, targeting coding and agent tasks. The company says Beam's performance is comparable to GLM‑5.2 and is gradually approaching Qwen3.8‑Max on coding and agent tasks; it has 501 billion total parameters and activates 23 billion per task.
Axios reports that Reflection and several other Western companies plan to launch open-weight AI models this month. Such models support local deployment, which matters to industries with strict data security requirements for fine-tuning and inference. The report expects Reflection’s model initially to trail the most advanced closed models from U.S. competitors. A Reflection spokesperson declined to comment.
AstaBrief 8B turns a research question and retrieved literature excerpts into a cited report. It is available in Asta’s Fast mode, and its weights and training data have been released. Across Asta’s full pipeline, Fast mode averages 51.1 seconds per report, versus 178.5 seconds for Thinking mode. Most evaluations were completed in 2025 and have not been fully rerun against today’s frontier models.