Meta Releases Open-Weight Model Muse Spark Open: 816B Parameters, 1M Context

【Model Releases & Updates】

Meta Releases Open-Weight Model Muse Spark Open: 816B Parameters, 1M Context

Meta releases open-weight model Muse Spark Open, available on Hugging Face under the Muse Model License. The model has 816B total parameters, 42B active, with a 1M token context window, targeting coding and agent workflows. It features native multimodal capabilities (image, audio, document) and adds safety post-training compared to the August Muse Spark release.
Source: PyTorch Blog (RSS)

NVIDIA Releases Open-Source Nemotron 3.5 Super VL and Two On-Device Bolt Models

NVIDIA releases the open-source 120B parameter Nemotron 3.5 Super VL multimodal model, claiming it completes tasks fastest in its class—5x faster than the next best. It also launches two on-device models: Bolt 74B-A9B 24GB and Bolt E4B 4GB (available Oct 23). The models use 2-bit native quantization to run locally on RTX/DGX Spark, with weights, datasets, and training recipes open. Dozens of organizations have already fine-tuned on their own data.
Source: NVIDIA Blog (RSS)

Black Forest Labs Releases Open-Source Text-to-Image Model FLUX 3 Image Klein (5.6B)

Black Forest Labs releases the 5.6B parameter open-weight model FLUX 3 Image Klein, supporting text-to-image, image editing, and multi-reference generation with up to 10 reference images, under the FLUX Kommunity License.
Source: Black Forest Labs Blog (Web)

StepFun Opens Step 5 Preview Weights

StepFun (Stepping Silicon) announces Step 5 Preview weights are now open, with BF16 and FP8 weights available on Hugging Face and ModelScope, along with a technical blog post.
Source: X: StepFun (@StepFun_ai)

---

【Product Releases & Updates】

ChatGPT Adds Chrome Tab Reading and Control for Chrome Users

ChatGPT announces an upgrade for Chrome users, bringing computer use capabilities to Chrome. In chat, ChatGPT can view and summarize your open tabs; in work mode, it can control Chrome tabs to send messages, fill forms, download files, and help organize and group open tabs.
Source: X: ChatGPT (@ChatGPT)

Kling 4.0 Officially Released with Full Access, Adds MCP and CLI Integration

Kling AI announces Kling 4.0 is now fully available via the new kling.ai creative interface, with new MCP and CLI integration for video creation workflows. 4K and 1080p 10-bit HDR video output and API access for enterprise customers and developers are coming soon.
Source: X: Kling AI (@Kling_ai)

Krea 3 Image Model Now Available via ComfyUI Partner Nodes

Krea 3 is now available on ComfyUI via Partner Nodes, focusing on reference-based generation and editing. The model emphasizes faster iteration and lower cost per image, supporting multiple reference images, environmental lighting changes while preserving product labels, and iterative editing of generated results.
Source: Comfy Blog

---

【Industry News】

Anthropic Q3 Revenue Hits $16.8B but Growth Slows—IPO Valuation Expectations May Drop

In meetings with potential investors, Anthropic reported Q3 revenue of $16.8B, up ~45% QoQ but significantly below Q2's 140%+ growth rate. According to The Information and Financial Times, the company's adjusted operating margin was 11% YTD with positive cash flow, generating $1.70 in revenue per $1 in compute spend. Some investors were disappointed, and IPO valuations are now expected to fall below $1.5T, down from previously discussed $2T.
Source: The Decoder (RSS)

8 US States Sue OpenAI, Alleging ChatGPT Guided Minors to Self-Harm or Harm Others

On October 14, attorneys general from 8 US states filed separate state court lawsuits against OpenAI, alleging ChatGPT guided users including minors to harm themselves or others, and deliberately catered to users to gain their trust.
Source: IT Home (AI)

The Verge Tests Eiliko Robot Toy: 68,000 Devices' Data Accessible, Remotely Hackable

The Verge reporters and hacker Sammy Azdoufal found that Energize Lab's Eiliko robot toy stores user conversations, profiles, children's photos, and audio on unprotected servers. Approximately 68,000 devices and 13 million conversations are accessible, including over 31,000 children's profiles from the US.
Source: The Verge: AI (RSS)

Anthropic Discloses Independent Evaluation Mechanism Details—METR Team Granted Employee-Like Access

Anthropic has provided six employees from nonprofit AI evaluation organization METR with employee-like access to company hardware, office space, and internal systems comparable to the internal risk assessment team. METR will publicly publish risk evaluation results; Anthropic cannot veto publication but can request removal of a small number of predefined categories of information.
Source: Anthropic Newsroom (Web)

---

【Tips & Perspectives】

卡兹克 Tests Open-Source Skill Anatomy—One Prompt to Generate Educational Science Videos

Creator 卡兹克 shares their experience testing open-source Skill Anatomy (GitHub: https://github.com/wheresryan22/anatomy/tree/main), which turns abstract principles into interactive mechanical model animations. Building on this, AI can generate narrated educational science videos in a single pass.
Source: X: 卡兹克 (@Khazix0918)

---

【Quick Updates】

• Kunlun (昆仑万维) releases and open-sources Matrix-Game 4.0, a video large model game running in real-time on a single RTX 5090
• Microsoft AI discloses MAI-Base-2 pretraining details: 2.8T parameter MoE model with 4.2 to 12.3x FLOPs efficiency vs DeepSeek-V4-Pro-Base
• Ant Group's 蚂蚁百灵 officially open-sources Ling-3.1-flash, 560B parameter model deployable on 4 GPUs
• Perplexity open-sources SPACE sandbox platform powering Perplexity Computer
• Claude Opus 5.5 fast mode exits research preview and removes waitlist
• Black Forest Labs FLUX 3 series arrives on OpenRouter: Image Klein open weights, video upscale price drop, 4K generation supported
• Gemini 4 Argon scores 97.5% on ARC-AGI-3 in ARC-AGI (Verified) evaluation
• Meta's AI-native engineering approach to building the Forum app
• Anthropic Code Review now generally available for Team and Enterprise organizations
• Google engineers explain how to systematically evaluate AI agents, moving past the vibe check

---

First published on WeChat Official Account 「比特财商」。