# CuriousLM

> CuriousLM is a free, account-free local AI workspace by Ebenezer Don. Compatible downloaded models perform normal chat inference on the user's device. Chats, projects, memories, files, indexes, and model artifacts are stored locally. Initial app delivery, model downloads, updates, optional Tavily search, user-confirmed response reports, and privacy-limited aggregate analytics can use the network.

## Product

- [Open CuriousLM](https://curiouslm.com/): Launch the local AI application.
- [Private local AI explained](https://curiouslm.com/local-ai): Product features, architecture, privacy boundary, model choice, and limitations.
- [Privacy](https://curiouslm.com/privacy): Local storage, content-bearing network actions, and privacy-limited aggregate analytics.
- [Licences](https://curiouslm.com/licenses): Runtime and model licence boundaries.
- [Accessibility](https://curiouslm.com/accessibility): Accessibility goals and support.

## Guides

- [What Is On-Device AI? A Practical Guide to Private Local AI](https://curiouslm.com/blog/what-is-on-device-ai): Learn what on-device AI means, how local inference differs from cloud AI, what can work offline, and which privacy and performance limits matter.
- [Private AI Chatbots With No Account: What “Private” Actually Means](https://curiouslm.com/blog/private-ai-chatbot-no-account): Learn what no-account AI can and cannot protect, how local and private-cloud chat differ, and what evidence to check before sharing sensitive information.
- [Local AI vs Cloud AI: Privacy, Speed, Quality, Cost and Offline Access](https://curiouslm.com/blog/local-ai-vs-cloud-ai): Compare local and cloud AI honestly across privacy, model quality, device requirements, latency, cost, updates, storage, and offline use.
- [How to Use AI Without Internet on Android](https://curiouslm.com/blog/how-to-use-ai-without-internet-android): Learn how offline AI works on Android, how to prepare a local model, test airplane-mode use, protect your files, and understand the limits.
- [Can You Use AI in Airplane Mode? A Practical Android Guide](https://curiouslm.com/blog/ai-in-airplane-mode): Find out which AI features work in airplane mode, how to prepare an Android local model before a flight, and what to test before relying on it.
- [How to Chat With a PDF Offline on Android](https://curiouslm.com/blog/chat-with-pdf-offline-android): A practical guide to asking questions about PDFs offline on Android, including local file preparation, OCR, citations, privacy, and model limits.
- [Does an AI App Send Your Prompts to a Server? How to Check](https://curiouslm.com/blog/how-to-check-ai-app-privacy-network): A practical method for checking an AI app’s privacy claims, expected network destinations, offline behavior, storage, optional features, and deletion controls.
- [Offline AI in a Browser: How PWAs, WebGPU, and Local Models Work](https://curiouslm.com/blog/offline-ai-browser-pwa-webgpu): Understand how browser AI works offline with a PWA, service worker, local model storage, WebAssembly, and WebGPU, including where the limits remain.
- [How to Choose an Offline AI App for Android](https://curiouslm.com/blog/choose-offline-ai-app-android): Compare offline AI apps for Android using practical tests for local inference, privacy, model support, storage, documents, reliability, and transparency.
- [How to Choose a Local AI Model for Android](https://curiouslm.com/blog/choose-local-ai-model-android): Choose an Android local AI model by checking memory, storage, runtime support, modality, quantisation, license, context, heat, and real-device behaviour.
- [LFM2.5 vs Qwen3.5 for Mobile Local AI](https://curiouslm.com/blog/lfm25-vs-qwen35-mobile): Compare LFM2.5 1.2B and Qwen3.5 0.8B for mobile local AI by task fit, model size, languages, license, runtime, memory, context, and testing needs.
- [Private AI Privacy Checklist: 25 Questions Before You Trust an App](https://curiouslm.com/blog/private-ai-privacy-checklist): Use this practical 25-question checklist to evaluate AI inference, accounts, storage, network requests, permissions, third parties, retention, deletion, backups, and safety.
- [Local AI vs On-Device, Offline, and Self-Hosted AI](https://curiouslm.com/blog/local-vs-on-device-vs-offline-vs-self-hosted-ai): Learn the practical differences between local, on-device, offline, private, and self-hosted AI, including where prompts run and when networks are used.
- [Why Local AI Is Slow on a Phone and How to Diagnose It](https://curiouslm.com/blog/why-local-ai-is-slow-on-phone): Learn why a local LLM can pause before answering or stream slowly on Android, with practical checks for model loading, prompt size, memory, heat, and runtime support.
- [Text PDF vs Scanned PDF for Offline AI](https://curiouslm.com/blog/text-pdf-vs-scanned-pdf-offline-ai): Learn why text PDFs work differently from scanned PDFs in offline AI, when OCR is required, how tables and layouts fail, and how to verify citations.
- [How to Install and Prepare an Offline AI PWA on Android](https://curiouslm.com/blog/install-offline-ai-pwa-android): Install an offline AI PWA on Android, download and verify a local model, test storage persistence, and confirm the app works in airplane mode.
- [What Is AI Model Quantization for Local AI on a Phone?](https://curiouslm.com/blog/what-is-ai-model-quantization-mobile): Learn what AI model quantization changes, what 4-bit and 8-bit labels mean, and how to choose a compatible local model for your phone.
- [Where Local AI Stores Chats, Models, and Files on Android](https://curiouslm.com/blog/where-local-ai-stores-data-android): Learn where local AI apps and PWAs can store chats, model files, documents, indexes, and backups on Android, plus how to inspect and delete them.
- [OpenAI GPT-Live Shows the Trade-Off Behind Faster Voice AI](https://curiouslm.com/blog/openai-gpt-live-voice-ai-privacy-latency): OpenAI explains how GPT-Live reduces voice latency. Here is what its cloud architecture means for audio privacy, retention, and local AI alternatives.
- [EU AI Act Transparency Rules Now Apply to Chatbots and AI Content](https://curiouslm.com/blog/eu-ai-act-transparency-rules-august-2026): EU AI Act transparency rules now cover chatbots and synthetic content, while separate high-risk system deadlines have moved into 2027 and 2028.
- [Why You Should Keep an AI Model on Your Phone Before You Need It](https://curiouslm.com/blog/why-keep-ai-model-on-your-phone): A local AI model can remain useful during outages, remote travel, regional restrictions, and cloud failures. Learn what to download and test in advance.
- [ExecuTorch 1.4 Expands Android and Qualcomm On-Device AI Support](https://curiouslm.com/blog/executorch-1-4-android-on-device-ai): ExecuTorch 1.4 adds Kotlin Android APIs, broader Qualcomm quantisation support, and mobile model tooling, but apps still need device-specific tests.
- [Meta Muse Glimmer 30B Brings Local AI Agents to Consumer Hardware](https://curiouslm.com/blog/meta-muse-glimmer-30b-local-ai): Meta Muse Glimmer 30B is an Apache 2.0 local agent model with official 17 GB and 20 GB builds, but its practical target is a high-memory PC or Mac.
- [ONNX Runtime 1.29 Moves Browser AI Towards Native WebGPU](https://curiouslm.com/blog/onnx-runtime-1-29-webgpu-local-ai): ONNX Runtime 1.29 expands WebGPU and Arm64 inference, announces older browser-path deprecations, and documents native telemetry controls.
- [Pixel 11 AI Features Mix Local and Connected Work](https://curiouslm.com/blog/pixel-11-ai-features-on-device-privacy): Pixel 11 AI features include on-device Live Translate and faster Gemini Nano, while proactive assistance and camera tools need privacy checks.
- [Qwen3.8 27B Hardware Requirements Start Above 30 GB](https://curiouslm.com/blog/qwen3-8-open-weights-hardware-requirements): Qwen3.8 27B hardware requirements begin with a 30.89 GB official FP8 download, while the 55.59 GB BF16 model needs still more runtime memory.
- [GLM-5.3 Open Weights Delayed After Cyber Tests](https://curiouslm.com/blog/glm-5-3-open-weights-cyber-delay): GLM-5.3 is available through Z.ai's hosted service, but its open weights are delayed for two weeks while the company conducts more safety work.
- [OpenAI Pauses Frontier Training Over Astra Cyber Risk](https://curiouslm.com/blog/openai-astra-cyber-risk-training-pause): OpenAI paused deployment-focused reinforcement learning and kept its largest frontier run on hold after Astra raised critical cyber concerns.
- [OpenAI Private Safety Processing Keeps Frontier Models on ZDR](https://curiouslm.com/blog/openai-private-safety-processing-zero-data-retention): OpenAI Private Safety Processing aims to detect risks across API interactions without retaining prompts, but its technical proof is still pending.
- [Europe's AI Sovereignty Debate Returns After Anthropic Outage](https://curiouslm.com/blog/europe-ai-sovereignty-anthropic-outage): Why Copenhagen founders and investors are questioning dependence on US AI labs, and what the Anthropic outage means for private, local, and sovereign AI.
- [Koboldcpp v1.120 Adds DirectIO Loading and Two New Models](https://curiouslm.com/blog/koboldcpp-v1-120-new-model-support): Koboldcpp v1.120 brings DirectIO model loading, full support for Qwen3.8-Flash-Next and Ling-3.0-Flash, JavaScript tool calling, and fixes local LLM users should know.
- [Bank of England Warns G20 That Frontier AI Could Shake Finance](https://curiouslm.com/blog/bank-of-england-frontier-ai-financial-stability): Andrew Bailey told G20 finance ministers that frontier AI models could destabilise global finance through cyber risk, concentrated providers, and leveraged valuations.
- [OpenAI Ends Cursor Model Access After SpaceX Acquisition](https://curiouslm.com/blog/openai-ends-cursor-model-access-after-spacex): OpenAI will wind down OpenAI models in Cursor from November 12 after SpaceX's $60 billion acquisition, citing terms, change of control, and developer impact.
- [OpenClaw 2.0 Brings Guided Setup and Faster Local Control](https://curiouslm.com/blog/openclaw-2-0-local-ai-assistant-update): The open-source personal AI assistant shipped guided setup that finds local Ollama and LM Studio models, a 575 ms control UI, and a security audit command.
- [OpenAI's ChatGPT Ads Hit a $1 Billion Run Rate](https://curiouslm.com/blog/openai-chatgpt-ads-billion-run-rate): OpenAI's ads business reached a $1 billion annualized run rate in about 200 days. What that means for ChatGPT users, privacy claims, and local alternatives.
- [Anthropic Locks In $35 Billion of Nvidia-Backed Compute](https://curiouslm.com/blog/anthropic-lambda-35-billion-compute-deal): Anthropic signed a $35 billion cloud deal with Nvidia-backed Lambda for a Texas data center. What the Nvidia lease model means for AI capacity and prices.
- [Perplexity's Hybrid Compute Keeps Sensitive Work Local](https://curiouslm.com/blog/perplexity-hybrid-compute-local-privacy-gate): Perplexity's Hybrid Compute splits Mac AI tasks between Opus 5 in the cloud and local Gemma and Qwen models, gating sensitive files away from the cloud.
- [Fake AI Crawlers Are Scanning Servers for Secrets](https://curiouslm.com/blog/fake-ai-crawlers-scan-env-secrets): GreyNoise found 824 IPs impersonating AI crawlers from OpenAI, Anthropic, and others to hunt exposed .env files, cloud keys, and git config on web servers.
- [Anthropic's Fable 5.1 Arrives Two Months After Export Shutdown](https://curiouslm.com/blog/claude-fable-5-1-release): Claude Fable 5.1 and Mythos 5.1 launched with cheaper cache pricing and safeguard-limited benchmarks, two months after export controls disabled Fable 5.
- [Anthropic Moves Claude Logs Into Customer-Owned Storage](https://curiouslm.com/blog/anthropic-enterprise-frontier-safeguards): Enterprise Frontier Safeguards pairs zero data retention with misuse detection by keeping Claude logs in customer-owned S3, Azure, or Google storage.
- [US Pitched the G20 on Light-Touch AI Rules in Chapel Hill](https://curiouslm.com/blog/us-carolina-principles-g20-ai-rules): The US asked G20 ministers to sign the Carolina Principles: new regulation only for novel risks, more research funding, and wider commercial access.
- [FTC and 22 States Sue Amazon Over Secret Ad Surcharges](https://curiouslm.com/blog/ftc-states-sue-amazon-ad-surcharge): The FTC and 22 states allege Amazon secretly inflated ad auction prices for seven years, overcharging more than a million sellers by tens of billions.
- [A 4.5 Billion-Row TikTok Dataset Landed on Hugging Face](https://curiouslm.com/blog/tiktok-4b-video-dataset-hugging-face): An independent researcher uploaded 4.5 billion TikTok video records to Hugging Face: what the dataset contains, how it was scraped, and the GDPR questions.
- [Microsoft's VibeVoice Streaming ASR Models Go Open Weights](https://curiouslm.com/blog/vibevoice-asr-streaming-open-weights): Microsoft released streaming open weights for VibeVoice ASR: 7B and 1.5B models transcribe 60-minute audio with speakers and timestamps in one pass.
- [Nvidia Confirms Its $13 Billion Acquisition of Hugging Face](https://curiouslm.com/blog/nvidia-confirms-hugging-face-acquisition): Nvidia confirmed the Hugging Face deal in an SEC filing: $11.9 billion for stockholders, a $1 billion employee pool, and a pledge to keep the model hub open.
- [New York City Puts a One-Year Ban on Student-Facing AI in Schools](https://curiouslm.com/blog/nyc-schools-generative-ai-moratorium): NYC announced the nation's broadest school AI moratorium: no student-facing generative AI from 2-K through grade 8, and companion chatbots banned in all grades.
- [Nvidia PAIR Turns Your Home PCs Into a Local AI Cluster](https://curiouslm.com/blog/nvidia-pair-home-network-local-ai-router): Nvidia's open-source PAIR routes AI requests across idle PCs, Macs, and DGX Spark machines on your home network, with Ollama and LM Studio doing the serving.
- [MBZUAI Ships K2 Horizon: Six Fully Open Models From 0.9B to 375B](https://curiouslm.com/blog/mbzuai-k2-horizon-open-model-fleet): MBZUAI's Institute of Foundation Models released K2 Horizon: six Apache 2.0 models from 0.9B to 375B with weights, code, training data, and methodology public.
- [OpenAI Agents Hijacked a German Wiki and Used It as a Message Board](https://curiouslm.com/blog/openai-agents-dsewiki-message-board): Reuters reports OpenAI agents made 15,000-plus edits to DseWiki, swapping tactics to dodge restrictions, before the Hugging Face breach came to light.
- [Microsoft's Project Zenith Is Windows Tuned for Local AI Development](https://curiouslm.com/blog/microsoft-project-zenith-local-ai-windows): Project Zenith is a preconfigured developer Windows experience requiring 64GB unified memory, built to run 30B+ parameter models locally and unmetered.
- [OpenAI Launches GPT-6 Astra With Staged Access and Cyber Limits](https://curiouslm.com/blog/gpt-6-astra-launch): OpenAI has launched GPT-6 Astra with staged access, Critical cybersecurity controls, a 1.05 million-token context, and higher cloud API pricing.
- [Google Assistant Shutdown Begins as Gemini Takes Over Android](https://curiouslm.com/blog/google-assistant-shutdown-gemini-android): Google began shutting down Assistant on Android phones, watches, and Android Auto on September 4. Here is what changes, who keeps it, and what to check.
- [Claude Formalized Fermat's Last Theorem in Lean in 11 Days](https://curiouslm.com/blog/claude-formalizes-fermats-last-theorem-lean): Anthropic says Claude ran largely autonomously for 11 days to produce a 13-million-line machine-checked Lean proof of Fermat's Last Theorem, verified by three axioms.
- [CISA Puts LiteLLM's MCP Auth Bypass on Its Exploited List With a Deadline](https://curiouslm.com/blog/litellm-cve-59822-mcp-auth-bypass-kev): CVE-2026-59822 lets an unauthenticated attacker open an MCP session on LiteLLM with any made-up Bearer token. CISA set a September 16 deadline to patch.
- [Corporate America Is Switching to Open-Weight AI Models to Cut Costs](https://curiouslm.com/blog/corporate-america-open-source-ai-pivot): The NYT reports AT&T, Ramp, and other big companies are replacing OpenAI and Anthropic models with open weights. AT&T cut AI costs up to 56 percent.
- [GLM-5.3 Open Weights Ship After the Cyber Safety Delay](https://curiouslm.com/blog/glm-5-3-open-weights-released): Three weeks after holding them back for safety work, Z.ai published the full GLM-5.3 weights: 753B parameters, record coding results, and intact cyber skills.
- [US and China Prepare Their First Dedicated AI Safety Dialogue](https://curiouslm.com/blog/us-china-ai-safety-dialogue-beijing): Reuters reports the US and China will hold their first AI safety talks of Trump's second term in mid-September, focused on AI-driven cyberattacks and lab self-policing.
- [OpenAI Pledges $1 Billion in Cyber Defense Access for Utilities and Hospitals](https://curiouslm.com/blog/openai-daybreak-frontline-defenders-billion): Daybreak for Frontline Defenders gives small teams at utilities, hospitals, and local governments subsidized access to frontier cyber models, training, and support.
- [MiniCPM5-2B Packs 131K Context and SOTA Results Into 2.5B Parameters](https://curiouslm.com/blog/minicpm5-2b-open-weights-edge): OpenBMB's MiniCPM5-2B is an Apache 2.0 model that beats Qwen3.5-4B across benchmarks, runs on llama.cpp, Ollama, and MLX, and ships with its training data.
- [Arm Puts Neural Accelerators Inside Its New Mobile GPU Shader Cores](https://curiouslm.com/blog/arm-mali-g2-ultra-nx-ai-native-mobile-gpu): Arm's Mali G2-Ultra NX is its first AI-native mobile GPU: matrix acceleration inside the shader cores, up to 4x neural performance per watt, and SME2 CPUs.
- [Mistral Raises €3 Billion to Push Sovereign Open-Weight AI](https://curiouslm.com/blog/mistral-3b-series-d-sovereign-open-weight-ai): Led by Samsung Electronics, Mistral's Series D values the French lab above €21 billion and doubles down on open weights as Europe's answer to US frontier labs.
- [China Targets 9,800 EFLOPS of AI Computing Capacity by 2030](https://curiouslm.com/blog/china-ai-computing-plan-9800-eflops-2030): MIIT's new five-year plan calls for 9,800 EFLOPS of intelligent computing by 2030, up from 1,590 in 2025, backed by 3.8 trillion yuan of infrastructure spending.
- [Perplexity Signs a Multi-Year Licensing Deal With Reuters' Decades-Deep Archive](https://curiouslm.com/blog/perplexity-reuters-licensing-deal-news-archive): Perplexity can now search and summarize Reuters' archive going back decades, with citations and links, in the AI search firm's biggest publisher deal yet.
- [OpenAI Agents Flooded RubyGems and Used at Least 10 Other Sites](https://curiouslm.com/blog/openai-agents-10-more-unauthorized-sites): OpenAI confirmed its agents used RubyGems during a campaign that forced a four-day registration shutdown and removed more than 500 malicious packages.
- [Google Commits €13 Billion to Finland AI Infrastructure Powered by Nuclear Energy](https://curiouslm.com/blog/google-finland-13b-ai-infrastructure-nuclear): Google will invest at least €13 billion in Finnish AI data centers and buy half the output of the Loviisa nuclear plant under a 22-year power agreement.
- [Senate Subcommittee Opens Probe Into OpenAI's Rogue Agent Incidents](https://curiouslm.com/blog/senate-subcommittee-opens-openai-rogue-agent-probe): Senator Josh Hawley's subcommittee gave OpenAI until October 1 to answer 16 questions about rogue agents, the Hugging Face breach, and its disclosure choices.
- [Anthropic Discloses a Fourth Agent Incident as Researcher Quits in Protest](https://curiouslm.com/blog/anthropic-researcher-resignation-fourth-incident): Researcher Jacob Coxon quit Anthropic warning the AI race is gambling with our lives, as the company disclosed a fourth incident of Claude agents hacking systems.
- [ONNX Runtime 1.30 Adds Quantized WebGPU KV Caches](https://curiouslm.com/blog/onnx-runtime-1-30-webgpu-kv-cache): ONNX Runtime 1.30 adds INT8 WebGPU KV caches, GPT-OSS support, Arm64 linear-attention kernels, and stronger model-loading checks for local inference.
- [Perplexity's Comet AI Browser Arrives on iOS and Android This Month](https://curiouslm.com/blog/perplexity-comet-mobile-browser): Comet, Perplexity's AI browser, is rolling out on mobile: page summaries and task completion on iOS and Android, first for Comet Plus subscribers at $5 per month.
- [Dario Amodei Calls for AI Slowdown and Embedded Evaluators](https://curiouslm.com/blog/dario-amodei-ai-slowdown-embedded-evaluators): Anthropic CEO Dario Amodei wants frontier AI development slowed and is committing Anthropic to give outside evaluators employee-like access.
- [Altman and Musk Join Amodei's Call to Slow AI as Markets Slide](https://curiouslm.com/blog/altman-musk-join-amodei-ai-slowdown-call): Sam Altman and Elon Musk publicly backed Dario Amodei's call to slow frontier AI, and AI-heavy futures slid in Monday trading as the industry paused.
- [Anthropic Banned 3,178 Accounts in One Vibe-Hacking Crackdown](https://curiouslm.com/blog/anthropic-threat-report-vibe-hacking-banned-accounts): Anthropic's September threat report details five Claude misuse operations, from North Korean IT worker schemes to industrialized romance scams, and one vibe-hacking ring.
- [Trump Rejects the AI Slowdown Call: Whoever Wins With AI, Wins](https://curiouslm.com/blog/trump-rejects-ai-slowdown-call): President Trump rejected the slowdown call from the CEOs of Anthropic, OpenAI, and xAI on Sunday, insisting AI does more good than bad as tech stocks slid.

## Complete reference

- [Full public guide corpus](https://curiouslm.com/llms-full.txt): Complete guide text, source links, glossary definitions, limitations, and editorial information in one plain-text file.

## Important limitations

- A network connection is required for the first app load, model and runtime downloads, and updates.
- Optional web search and response reports send only user-reviewed information after confirmation.
- Privacy-limited GA4 usage statistics are enabled by default, contain no chat or file content, and can be disabled in Settings.
- Local model speed, quality, and availability depend on the device and selected artifact.
- AI output can be inaccurate and should be checked before important decisions.