Follow

Microsoft launches MAI-Image-2.7-Preview for image editing

Microsoft has launched the MAI-Image-2.7-Preview image editing model, which users can try on Arena, Artificial Analysis and MAI Playground. It currently edits existing images using text instructions; text-to-image generation is not yet available. Microsoft says it improved over MAI-Image-2.6 across all 10 use cases and seven editing categories in an Artificial Analysis evaluation.

LFM3 models released with open weights for text, vision and audio

The six LFM3 models are available with open weights on Hugging Face. They cover text, vision, audio, multimodal and decision tasks, and each can be fine-tuned on a single GPU. The developer says LFM3-1.7B outperforms its predecessor, LFM2.5-2.6B, with 35% fewer parameters. Its first omni model, LFM3-Omni-2B, accepts text, images, video and audio and produces text or speech.

Genie 4 research preview reaches trusted testers with real-time, audio-enabled world simulation

Genie 4 generates ambient sound and interactive dialogue within simulated worlds, while text prompts can change the weather, camera or scene in real time. The research preview is currently available to trusted testers; access for Google AI Ultra subscribers and selected developer partners is still forthcoming. Scenes can remain coherent over short sessions, but visual details and object permanence may drift over longer runs or after multiple edits.

4 more sources reported this

• IT之家 科技新闻
• Rohan Paul (@rohanpaul_ai)
• TechCrunch AI
• The Verge AI

OpenAI releases GPT-6.1 Terra for desktop tasks and enterprise automation

GPT-6.1 Terra is available to Plus, Pro, Enterprise and Edu users and through the API. It targets tasks such as invoice processing and interpreting technical drawings. On the cited benchmark at medium reasoning effort, it scored 15 percentage points above its predecessor at about one-third the cost per task. OpenAI still recommends other models for the hardest coding tasks.

1 more source reported this

• Tibo (@thsottiaux)

Moonshot releases Kimi K3.1 Pro after reducing its exploit capability

Moonshot says removing cyber-related training data lowered Kimi K3.1 Pro’s pass rate on its internal exploit test from 90.5% for the unfiltered version to 81.0%, still above K3’s 33.3%. The model and a Highspeed variant running 2–3 times faster are live in Kimi apps and the API; weights are due by November 4, 2026.

2 more sources reported this · 2 related updates

• IT之家 科技新闻
• MarkTechPost AI 研究报道

• NVIDIA · Change苹果盘中总市值暂时超越英伟达，使英伟达暂时失去全球上市公司市值榜首位置。
• Jensen Huang · Statement黄仁勋表示对英伟达与长鑫存储合作持开放态度，并称中国在基础设施技术和制造业方面能力很强，应认真对待。

Mosi and OpenMOSS release MOSS-Image with open weights

The 5B-parameter image model offers weights and an API. Developers can use JSON Layout to place text and objects. Its 96.3% score on the LLTB benchmark measures text coverage, not character-by-character accuracy.

MiniMax releases M3.1 Flash model weights, with 428B total parameters

MiniMax announced the release of M3.1 Flash model weights. The model has 428B total parameters and 23B active parameters. The company says its capabilities improve across the board over M3 and that sparse attention and quantization-aware training improve inference efficiency. After its October 19 launch, real users processed more than 100 trillion tokens within 72 hours.

2 more sources reported this

• IT之家 科技新闻
• ginobefun (@hongming731)

OlmoEarth releases v1.3 satellite model and browser-based map explorer

OlmoEarth v1.3 includes precomputed satellite embeddings and a browser-based tool for exploring landscape changes and finding similar places without an account. Global coverage currently spans 2023–2025; coverage for 2017–2022 is limited to select countries. The release says its 128-dimensional embeddings scored highest in 20 of 27 evaluations against two other models.

iFLYTEK releases SparkAudio 1.0 across its input method and other products

iFLYTEK has released SparkAudio 1.0, a speech foundation model combining machine translation, speech recognition and semantic understanding. It is available in iFLYTEK Input Method 16.0, iFLYTEK Tingjian, iFLYTEK AI Office Notebook and iFLYTEK Translator, with Chinese-English switching and real-time speaker identification. The company also released the iFLYTEK AI Recording Ring, which it says can transcribe two hours of audio in two minutes.

NVIDIA Adds StormScope EU as European Organizations Apply Earth-2 Models

NVIDIA has added StormScope EU for short-term forecasting over Europe, predicting cloud conditions at 10-minute intervals. AXA is studying cyclone risk, while Vaisala Xweather is working on short-term warnings. University of Manchester researchers adapted a model for U.K. air-quality modeling; TotalEnergies is evaluating solar-radiation nowcasts, and Space42 developed a model to estimate surface heights from radar imagery.

1 related update

• NVIDIA · Other英伟达在GTC Berlin展示了与Earthmover合作开发的交互式天气演示，其中运行了StormScope EU实时预测流程。

DeepSeek releases V4.1-Flash-1017 across web, app and API

DeepSeek V4.1-Flash-1017 is now available on the web, app and API, with API access and pricing unchanged. DeepSeek says it improves on the previous version while reducing task time and output length, and is better able to generate runnable projects from limited prompts and reference materials. Model weights are also available to download.

Reflection releases Beam with open weights; Ollama offers free cloud access for a limited time

Reflection's first open-weight model, Beam, is available, with Ollama offering free cloud access for a limited time without requiring a model download or local inference management. Beam targets coding, reasoning, and multi-step agent tasks, supports a 1 million-token context window, and runs on an Ollama endpoint that the company says retains no data.

Ant Ling opens Ling-3.1-flash with support for up to 1M context

Ant Ling announced the release of Ling-3.1-flash as an open-source model. It supports up to 1M context and has improved JSON Schema support for tool calls. In the specified Hopper GPU test configuration, quantization reduced weight storage from about 534GB to 292GB and the serving setup from eight GPUs to four; actual requirements still vary with context length, concurrency and service configuration.

1 more source reported this

• IT之家 科技新闻

NVIDIA releases Nemotron 3.5 Super VL and makes Bolt 74B available

NVIDIA has released Nemotron 3.5 Super VL for multistep agent tasks, claiming up to five times the single-GPU task throughput of other open models of similar size. Super VL and the Bolt 74B model, compressed to 24GB, are available; Bolt E4B 4GB, intended for uses including local voice assistants, is planned for October 23.

2 more sources reported this

• TechCrunch AI
• NVIDIA 开发者博客

Step 5 Preview weights released as four chipmakers announce day-one support

Step 5 Preview’s BF16 and FP8 weights are available on Hugging Face and ModelScope. Developers can deploy an OpenAI-compatible API service with vLLM or SGLang, or run offline inference with Transformers. Huawei, Biren Technology, MetaX and Moore Threads subsequently announced day-one support.

1 more source reported this

• IT之家 科技新闻

LFM2.5-24B-A2B open weights available for local reasoning agents

Open weights for LFM2.5-24B-A2B and its base version are now available on Hugging Face for agents that reason and call tools on-device. Its context window grows from 32K to 131K tokens over its predecessor. In the publisher's tests, it scored 48.09 on LiveClaw, while larger models retained an edge on some knowledge and coding benchmarks.

1 more source reported this

• MarkTechPost AI 研究报道

Gemini 4 Argon Becomes Available With Output of Up to About 1 Million Tokens

Gemini 4 Argon is available to developers and businesses worldwide through AI Studio and the Gemini platform API; eligible Google AI Ultra and Pro subscribers can also use it in supported products. Input, reasoning and output share a context window of about 1 million tokens. Introductory prices are $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the promotion ends on November 30.

1 more source reported this

• IT之家 科技新闻

Meta releases its first frontier models, Muse Wonder and Muse Wonder Full

Meta says the two models are larger than Muse Spark and perform better on complex tasks such as coding; in Meta’s tests, their benchmark scores were comparable to Fable 5.1 and GPT-6 Astra. Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens. Wonder Full is limited to Muse Code subscribers and is not offered through the API.

3 more sources reported this

• Latent Space AI
• IT之家 科技新闻
• Simon Willison AI 实践

Qwen-Image-2.1 Pro and Turbo APIs launch; Turbo weights released

Qwen-Image-2.1 Pro and Turbo APIs are now available, and Turbo weights have been released. The announcement says the image generation and editing tools support fast inference in eight steps, editing with up to 10 images, and native transparent layers for integration into applications.

1 more source reported this

• IT之家 科技新闻

Same event; featured as “千问 AI 平台上线 Qwen-Image-2.1 Pro 与 Turbo API，开放 Turbo 权重”

Shengshu Technology releases Vidu Q4 Preview with 4K video output

Vidu Q4 Preview is available as a preview. It can generate video from an image and text, or use up to 15 reference images and three audio clips. A single generation can run for up to 16 seconds, with 2K or 4K output; introductory pricing starts at 0.09 yuan per second.