AI Tool Watch

Advanced voice is now available on every plan, including free—OpenAI's GPT-Live replaced the old Advanced Voice Mode as the default, with GPT-Live-1 mini for free accounts and GPT-Live-1 for paid. Free users get it on web and mobile; voice in the macOS and Windows desktop apps is paid-only and returned in July 2026 after being retired in January.ℹ️GPT-Live replaced Advanced Voice Mode (July 2026): OpenAI's GPT-Live is now the default ChatGPT voice experience. Per the GPT-Live system card, GPT-Live-1 is the default voice model for paid users and GPT-Live-1 mini for free users. The legacy Advanced Voice Mode remains reachable only as a fallback for features GPT-Live does not yet support (video and screen sharing) on web and mobile. Usage limits: OpenAI publishes no per-tier numeric caps. The model release notes state only that "usage of advanced Voice (audio inputs and outputs) by Plus and Team users is limited on a daily basis." Limits are dynamic and surfaced in-app. Specific minute figures circulating in third-party guides are not OpenAI-confirmed. macOS desktop voice — retired then restored: • January 15, 2026 — OpenAI retired all voice functionality from the ChatGPT macOS desktop app, to "focus on more unified and improved voice experiences across our apps." Voice remained on web, iOS, Android, and the Windows app. • July 23, 2026 — ChatGPT Voice returned to the desktop app on both macOS and Windows, powered by GPT-Live, rolling out globally to Plus, Pro, Business, Edu, and Enterprise plans. It integrates with ChatGPT Work and Codex for voice-directed agent control. Free-tier users still have no desktop voice access and must use chatgpt.com or the mobile apps.

ChatGPT's Agent Mode lets it browse the web and take actions on your behalf. It's available on Plus ($20/mo) and higher—the new $8 Go plan does NOT include it. Plus gives you 40 agent tasks per month; Pro gives you 400. Now available in EEA—basic agent tasks rolled out to Europe in late July 2025; some Connector features remain restricted.

Atlas is OpenAI's AI-native browser. Right now it's macOS only—Windows is expected in the next couple months. You can use the browser on the free plan, but Agent Mode inside Atlas requires Plus or higher.

Canvas is being sunset. OpenAI removed the split-screen Canvas interface from its GPT-5.5 flagship models (Instant and Thinking), replacing it with inline writing blocks and code blocks in the chat thread. Canvas remains available by default in GPT-4o for all users including free, and paid users can still reach it through legacy models—but that legacy path ends when OpenAI o3 retires on August 26, 2026.

ChatGPT Search provides real-time web results with cited sources—news, weather, stocks, and more. Available on all plans including free and even logged-out users. Powered by Bing and partner publishers.

Codex is ChatGPT's coding agent that can write, run, and debug code autonomously. Available on every plan including Free, but usage is quota-gated—Free and Go get limited runs, while Plus ($20/mo) and above get the usable allowances. Available via web, desktop apps, and CLI (`npm i -g @openai/codex`) on macOS, Linux, and Windows (WSL).

Codex Security is OpenAI's AI security agent—it connects to your GitHub repos, builds a codebase-specific threat model, validates vulnerabilities in sandboxed environments, and generates fix PRs. Requires Pro ($200/mo), Enterprise, Business, or Edu—not available on Plus or lower. Launched March 6, 2026 as a research preview with free usage for the first month. As of July 29, 2026, an open-source Codex Security CLI (`npm install @openai/codex-security`) is also available for scanning repositories from the command line, in beta and requiring access.

Custom GPTs let you create specialized AI assistants with custom instructions, knowledge files, and Actions for external API integrations. Actions replaced the old plugin system in April 2024. Anyone can use public GPTs, but creating your own with Actions requires Plus ($20/mo) or higher.

ChatGPT image generation now uses ChatGPT Images 2.5 (rolled out September 8, 2026 across desktop, mobile, and web), which replaced the earlier GPT Image 1.5 default. Available on all plans including free with tiered limits. The DALL-E 3 API is deprecated May 12, 2026—developers should migrate to `gpt-image-2`, `gpt-image-1`, or `gpt-image-1-mini`. Note: `gpt-image-1` and `gpt-image-1.5` are themselves now slated for API retirement (Oct 23, 2026 and Dec 1, 2026 respectively), in favor of `gpt-image-2.5-sunburst`/`gpt-image-2.5-flare`.

Deep Research lets ChatGPT spend minutes researching a topic thoroughly. Now available on all plans including free—free users get 5 lightweight queries/month (powered by o4-mini). Plus gets 10 full + 15 lightweight per month; Pro gets 125 + 125.

Memory lets ChatGPT remember your preferences and context across conversations. All plans get saved memories; Plus/Pro also get full chat history reference for richer personalization.

Projects let you group related chats, files, and custom instructions into workspaces. Available on all plans including free—organize your work across topics with project-specific context.

Sora generates videos from text prompts—up to 25 seconds at 1080p. Requires Plus ($20/mo) or Pro ($200/mo). Plus gets 1,000 credits/month (~50 short videos); Pro gets 10,000 credits plus unlimited relaxed-mode generations. Not available in Europe.

GPT-4o understands images natively—upload a photo, screenshot, chart, or diagram and ChatGPT will analyze it. Available on all plans including free, subject to your plan's message quota. Works across web, desktop, iOS, and Android. Also accessible via the OpenAI API using the `gpt-4o` model.

ChatGPT's core text conversation is available on all plans including free at chatgpt.com, with apps for iOS, Android, Windows, and macOS. The model lineup has moved on twice since GPT-5.4 (March 2026): GPT-5.5 became the default generation, and GPT-5.5 itself is scheduled to retire from ChatGPT, ChatGPT Work, and Codex on all plans on October 14, 2026. OpenAI is actively rolling out its replacement now—GPT-6 Sol for Plus/Pro/Business/Enterprise/Edu and GPT-6 Luna for Free/Go (from September 22, 2026)—on top of GPT-6 Astra, OpenAI's most capable model, which reached Pro/Enterprise/Business-Premium plans and the API on September 4, 2026.

Agent Builder lets you create custom AI agents using natural language—no coding required. Requires M365 Copilot ($30/user/mo) for full access. As of March 2026, agents can call other agents as tools (multi-agent coordination), and new agent surfaces include Outlook, OneDrive, and Teams communities.

Copilot Connectors let M365 Copilot access external data from services like Salesforce, ServiceNow, Box, and more. Requires M365 Copilot ($30/user/mo), and is not available on the free tier or any Microsoft 365 consumer plan. Federated connectors use MCP to pull live data from services such as Notion, HubSpot, Linear, Canva, Google Contacts, and Google Calendar; Microsoft-published ones are on by default unless an admin turns them off, and write actions begin rolling out in early October 2026.

Copilot in Word, Excel, PowerPoint and Outlook now comes with any paid Microsoft 365 consumer plan, starting at $9.99/mo for Personal, or with a Microsoft 365 Copilot commercial license ($30/user/mo). Free Copilot still doesn't include Office integration. Two catches: the AI benefits go to the subscription owner only, even on Family, and usage draws on monthly AI credits. Copilot Pro is closed to new customers; existing subscribers can continue or switch to Microsoft 365 Premium. This table covers plans available to new customers; legacy Copilot Pro feature entitlements need separate review.

Copilot Vision lets Copilot watch your screen, a shared app, or your phone camera and talk you through what it sees. It runs inside a Copilot Voice session, so you share a screen and then ask out loud. It is no longer free: you now need Microsoft 365 Personal, Family, or Premium, capped at 10 minutes a day on Personal and Family and 15 minutes a day on Premium. It reaches well beyond Edge now, covering Windows, macOS, and the mobile apps. This table covers plans available to new customers; legacy Copilot Pro feature entitlements need separate review.

Copilot Voice is free once you sign in, subject to capacity. Subscribers get both priority and a published daily allowance: 30 minutes a day on Microsoft 365 Personal and Family, 60 minutes a day on Premium. Copilot Pro is closed to new customers; existing subscribers can continue or switch to Microsoft 365 Premium. Speak to Copilot on Windows, Mac, mobile, and the web. This table covers plans available to new customers; legacy Copilot Pro feature entitlements need separate review.

Copilot no longer sells access to a named model tier, and Microsoft's own plan documentation never mentions one. What a Microsoft 365 subscription buys is usage, not a better model: higher limits, priority access when capacity is tight, and on Premium the advanced AI features and the Analyst and Researcher agents. Reporting through 2026 describes Copilot routing across newer OpenAI models alongside Microsoft's own MAI models, with a model picker on some surfaces. Any guide still framing this as free-versus-Pro GPT-4 access is out of date. This table covers plans available to new customers; legacy Copilot Pro feature entitlements need separate review.

Image generation with Designer is available on every plan including free, though you must sign in. Paid plans meter it with AI credits: 60 a month on Microsoft 365 Personal and Family, and usage beyond the standard credit limit on Premium. Microsoft doesn't publish a monthly figure for free accounts. Powered by GPT-Image-1.5. The Designer bot in Teams was retired in February 2026, so use Copilot image generation instead. This table covers plans available to new customers; legacy Copilot Pro feature entitlements need separate review.

Microsoft Copilot's core text chat is free at copilot.microsoft.com, with no account required for basic access. Sign in with a Microsoft account for chat history, longer conversations, image creation, and voice. A paid Microsoft 365 plan buys higher usage, not a different chatbot, and Premium adds the Analyst and Researcher agents. Native apps available on Windows, macOS, iOS, and Android. This table covers plans available to new customers; legacy Copilot Pro feature entitlements need separate review.

AI Studio is Google's developer playground for Gemini models—test prompts, get API keys, and prototype before building. Completely free with generous rate limits.

Audio Overview transforms your content into AI-generated podcast-style discussions. Free users get 3 per day; Google AI Plus gets more, Google AI Pro gets 20, and Google AI Ultra has the highest limits. Originally from NotebookLM, now integrated into Gemini.

Gemini Canvas is an interactive workspace for creating and refining writing and code projects. Available on all plans including free—free users get limited prompts per day, Google AI Plus doubles free usage, and Google AI Pro unlocks a 1 million token context window.

Gemini's Deep Research creates comprehensive reports on complex topics. Now available on the free tier (~5 reports/month via Gemini Flash); Google AI Plus ($4.99/mo) doubles free usage; Google AI Pro ($19.99/mo) gets full access with Gemini 3.0 Pro for deeper research.

Gemini Connected Apps (formerly Extensions) connect Gemini to Google services like Gmail, Calendar, YouTube, Maps, and more. Available on all plans including free. The Gemini CLI also supports extensions via MCP for developers.

Gemini Advanced was Google's name for premium model access; Google's plans pages no longer use it and its help article is gone. Google AI Pro ($19.99/mo) unlocks Gemini 3.1 Pro with 4x free usage limits; Google AI Ultra (from $99.99/mo) adds the highest limits and Deep Think. Free users now get varying access to 3.1 Pro rather than none. Marked deprecated — see the Chat record for current model access.

Gemini integrates directly into Google Docs, Sheets, Slides, and Drive. March 2026 beta rollout adds 'Help me create' in Docs (instant drafts from files/emails), 'Fill with Gemini' in Sheets (auto-populate tables), new slide creation in Slides, and AI Overviews + 'Ask Gemini' in Drive. Google AI Plus ($4.99/mo) covers Gmail and Vids; Google AI Pro ($19.99/mo) adds Docs; Google AI Ultra (from $99.99/mo) has the highest limits; or the Workspace add-on for business accounts. Drive features are US-only for now; full deck generation in Slides is coming soon.

Gemini Live is Google's real-time voice mode in the Gemini mobile app, with interruption and back-and-forth built in. Included on the free plan. Google AI Plus ($4.99/mo), Google AI Pro ($19.99/mo) and Google AI Ultra raise usage limits (2x, 4x, up to 20x); Daily Brief inside Live starts at Google AI Plus and Gemini Spark inside Live starts at Google AI Pro. You can also share your camera or screen during a Live chat so Gemini can see what you see. Android and iOS only — Google's help pages say it isn't available in the Gemini web app.

Gems are custom AI experts you create for specific tasks. Now free for all users—create premade or custom Gems at no cost. Google AI Plus doubles free usage; Google AI Pro gets higher limits and sharing features.

Gemini can generate images using Imagen 4. Now free for all users—free tier gets ~10-20 images/day, Google AI Plus doubles that, and Google AI Pro gets ~100-150/day. Also available via the Gemini API.

Nano Banana Pro is Gemini's native image generation and editing model, built on Gemini 3 Pro. It's distinct from Imagen—optimized for conversational editing, character consistency, and multi-image fusion. Supports 1K-4K resolutions with SynthID watermarks. Full access requires Google AI Pro ($19.99/mo)—free and Google AI Plus users get Nano Banana 2 as the default in the Gemini app, with only limited Nano Banana Pro access through Google Flow. Access Nano Banana Pro via the three-dot model menu.

NotebookLM lets you upload documents and have AI conversations grounded in your sources—it even generates podcast-style audio summaries. Free to use with generous limits; Google AI Plus adds more notebooks and Audio Overviews, Google AI Pro raises source limits, and Google AI Ultra has the highest. Google's plans page now lists it as

Project Astra is Google DeepMind's research prototype for a universal assistant. Its camera and screen sharing now ship in Gemini Live, which is free for everyone on iOS and Android. The prototype itself remains waitlist and trusted-tester only. Marked deprecated as a product record — see Gemini Live for current availability.

Veo is Google's video generation model. Veo 3.1 generates 4K video up to 8 seconds with native audio—including dialogue, sound effects, and ambient noise. Starts at Google AI Plus ($4.99/mo, 200 Flow credits); Google AI Pro ($19.99/mo, 1,000 credits) and Google AI Ultra (from $99.99/mo, 10,000–25,000 credits) for the highest limits. Available in the Gemini app, Flow, and via API. Supports text-to-video and image-to-video with up to 3 reference images—though image-to-video is blocked in the EEA, UK, and Switzerland.

Gemini can now use screen automation (Google's own name for this beta) to complete multi-step tasks inside third-party Android apps—booking rides, flights and accommodations, ordering food and groceries, buying event tickets, and (on newer devices) buying/selling items—with the user able to view, take control of, or stop the task at any time. Available on Pixel 10, Pixel 10 Pro, Pixel 10 Pro XL, Samsung Galaxy S26 series, Z Flip 8, and Z Fold 8, in the US and Korea (Pixel 10 excluded from Korea). Still beta, English/Korean only, with select apps. No paid plan required.

Gemini's Past Chats feature lets it remember your preferences and context across conversations. Enable it in Settings & help → Personal Intelligence → Memory. Now free for all users—it rolled out to non-paying accounts in late February 2026. Requires a personal Google Account (18+) with Keep Activity on; not available on work/school/supervised accounts, in Gems, or in Live chats.

Gemini's core text Q&A is available for free at gemini.google.com to anyone with a Google account, in 230+ countries on web and 150+ countries on mobile. Free tier uses Gemini 3.6 Flash; Google AI Pro and Ultra unlock Gemini 3.1 Pro with higher limits. Gemini side panel now available in Chrome (Jan 2026) with Auto Browse for Google AI Pro/Ultra subscribers. A native Gemini app for Mac launched April 2026 (gemini.google/mac).

Artifacts are Claude's way of creating standalone content—a design, a deck, a document, a dashboard, a small interactive tool—in a separate panel. Available on Free, Pro, Max, Team, and Enterprise, and in Claude Code on any plan that includes it. Paid plans also get beta templates—Claude Design, Claude Slides, and Claude Docs—for polished, shareable output; on Enterprise these are off until an owner turns them on. Artifacts made before September 16, 2026 are 'legacy' artifacts and keep working, but new ones use the current experience.

Claude Code is Anthropic's agentic coding tool. It requires Pro subscription at minimum—that's $20/month. Available as a terminal CLI, a native VS Code extension, and a native JetBrains plugin (IntelliJ IDEA, PyCharm, Android Studio, WebStorm, PhpStorm, GoLand), plus Xcode integration via the Claude Agent SDK. It also now runs in the browser at claude.ai/code ('Claude Code on the web'), and the Claude mobile app (iOS/Android) lets you start, monitor, and steer cloud sessions or use Remote Control on a local session. A Chrome extension adds browser automation (reading console/DOM, testing, form-filling) to coding tasks on any paid plan.

Connectors let Claude access your apps and services—Google Drive, Notion, Slack, and more—browsable in the Connectors Directory. Free on all plans, including remote MCP connectors (Free is capped at 1 custom connector). Custom connectors via remote MCP server URLs are available on Claude, Cowork, and Claude Desktop. Browse available connectors at claude.ai/directory.

Cowork brings Claude Code's agentic capabilities to knowledge work beyond coding—it launched January 2026 and expanded to all paid plans in February. It's now available on web, Claude Desktop (macOS/Windows), and the Claude mobile apps, not just desktop, with the Chrome side panel rolling out from Max/Team to Pro. As of September 2026, Anthropic is merging Cowork and chat into a single 'Claude' experience ('Claude Cowork and chat are one Claude'), rolling out gradually starting with Pro and Max: there's no more mode to pick—any conversation can hand off a task the way Cowork used to.

Extended Thinking lets Claude reason through complex problems step-by-step. As of the Claude 5.5 family (Sonnet 5.5, Opus 5.5, Fable 5.1), thinking cannot be turned off—it's always on, at every effort level. Effort now has five levels: low/medium/high/xhigh/max, with 'Extra high' (xhigh) added for long-running coding and agentic tasks on Opus 4.7 and newer. Older models (Sonnet 4.6, Opus 4.6, Opus 4.7, Sonnet 5, Fable 5) still let you toggle thinking manually. Available on all plans including free (with usage limits); Pro and above get full unrestricted access.

MCP is an open standard for connecting Claude to external tools and data sources—databases, file systems, APIs. Available on all plans including free via the Claude Desktop app.

Projects let you organize conversations and documents into workspaces with persistent context. Available on all plans including free as of February 2026 (Free is capped at 5 projects). On paid plans, project knowledge automatically scales through Retrieval-Augmented Generation (RAG), expanding capacity up to 10x as content approaches context limits. A new version of projects (beta) is now rolling out too, starting with Claude Code: a project becomes one conversation that Claude splits into parallel cloud threads.

Skills are modular folders of instructions and resources that Claude can load on demand to perform specialized tasks—like creating Office documents or running workflows. As of a July 2026 update, both prebuilt and custom skills are available on Free, Pro, Max, Team, and Enterprise plans (custom skills need Cloud code execution turned on). The one piece still gated to paid plans is recording a skill—showing Claude a screen recording of a task so it drafts the skill for you—available on Pro, Max, and Team in Cowork on Claude for Mac only.ℹ️Prebuilt vs Custom Skills: • Prebuilt Skills: Created and maintained by Anthropic (Excel, Word, PowerPoint, PDF creation). Activate automatically when relevant. Available on all plans including free. • Custom Skills: User- or organization-created skills. Available on Free, Pro, Max, Team, and Enterprise (requires code execution enabled); Team/Enterprise can provision org-wide. Skill recording (build a skill from a screen recording) is Pro/Max/Team only, on Claude for Mac.

Claude can analyze images you upload—documents, screenshots, diagrams. Available on all plans including free.

Memory lets Claude remember your preferences and context across conversations—corrections, working style, recurring topics. Free on Free, Pro, and Max by default (Team/Enterprise: off by default, owner-enabled). Memory is shared between chat and Claude Cowork when Cowork runs in the cloud. A related but separate capability, searching your past chats with Claude (RAG-based), requires Pro, Max, Team, or Enterprise. Enable/manage both in Settings → Memory. You can also import memories from other AI tools using the memory import tool.

Claude's core text conversation is available on all plans including free at claude.ai, with apps for Windows, macOS, Linux, iOS, and Android. The free tier now runs on Claude Sonnet 5.5 (the Sonnet 4.x line is long superseded) with 5-hour session usage caps; Pro unlocks Opus access and 5x the usage of Free, Max adds a 5x or 20x-of-Pro tier. As of September 2026, Anthropic is merging Cowork and chat into one 'Claude' experience on Pro/Max first (see the Cowork Mode entry in this file), so the line between 'just chatting' and handing off a multi-step task is disappearing.

Perplexity Computer is a cloud-based agent that orchestrates 19+ AI models for multi-step workflows. Basic actions work on free; Pro ($20/mo) unlocks advanced capabilities. Max ($200/mo) adds full Computer access, Voice Mode, and Personal Computer—a dedicated Mac mini that runs 24/7 as your local agent. Enterprise gets Computer for Enterprise with Slack/Snowflake/Salesforce integrations. Four platform APIs (Search, Agent, Embeddings, Sandbox) are now available for developers.

Buy with Pro lets you research and purchase products directly in Perplexity—no ads, just AI recommendations. Now available on all plans via PayPal integration. Free shipping on all orders. Currently US only.

Collections let you organize research into projects. Available on all plans, with more features on Pro.

Perplexity's Comet browser is free for all users—available on Mac, Windows, Linux, Android, and now iPhone. Comet Enterprise (Mar 2026) adds managed deployment with admin controls, domain permissions, MDM support, audit logs, and CrowdStrike Falcon integration for enterprise security.

Enterprise App Connectors let you search across company files and connected apps like Google Drive, Confluence, and Slack. Enterprise Pro only ($40/user/mo)—not available on consumer plans.

Focus modes let you search specific sources—Academic, YouTube, Reddit, etc. Available on all plans including free.

Pro users can choose which AI model answers—GPT-5, Claude, Gemini, and others. Free users get Perplexity's default model only.

Pages let you turn research into shareable, formatted articles with AI-generated sections and citations. Available on all plans including free. Created pages are public and searchable.

Pro Search does deeper research with follow-up questions. Free users get 5-10 Pro searches per day; Pro subscription makes it unlimited.

Perplexity Voice allows hands-free search on mobile. Available on all tiers including free—just tap the microphone in the app.

Perplexity's core search-and-answer experience is free with unlimited basic queries at perplexity.ai, on iOS, Android, the web, a Chrome extension, and via the Comet desktop app. Free users get full access to web-grounded Q&A; Pro unlocks more powerful models and unlimited Pro searches.

DeepSearch is Grok's multi-step research mode. Premium+ or SuperGrok recommended for full access; basic Premium has limited DeepSearch.

Grok is xAI's chatbot, available through X (Twitter) or via grok.com. No official desktop apps exist—access is via web, iOS, or Android. Free users get very limited access; Premium ($8/mo) unlocks regular use. SuperGrok at $30/mo gets you the highest limits.

Grok Studio is a split-screen workspace for creating documents, running code, and building games—similar to Canvas or Artifacts. Available on all tiers including free at grok.com. Supports Python, JavaScript, TypeScript, C++, and HTML preview.

Grok Voice Mode allows natural voice conversations with Grok. Now free for all users on iOS, Android, and web (grok.com) with 11 voice modes. TTS API available for developers with 5 voice personalities. Supports attachments and photos in voice conversations (Mar 2026).

Grok has real-time access to X/Twitter posts—its unique advantage over other chatbots. Available on the free tier with lower limits; paid tiers (X Premium from $8/mo, SuperGrok $30/mo, SuperGrok Heavy $300/mo) get higher quotas and more capable models. Also accessible inside Tesla vehicles.

Think Mode lets Grok reason through complex problems step-by-step. Available on all paid tiers, with SuperGrok getting the longest thinking time.

Grok 2 understands images—upload photos, screenshots, or diagrams and Grok will analyze them. Requires a paid plan (Premium $8/mo or higher). Available on web and mobile.

Grok has persistent memory that carries user preferences and context across conversations. Available on all tiers including free. Memory is selective and automatic—Grok stores key facts it deems important. Users can turn memory off under Settings → Data Controls, but there is no itemized memory manager listing everything saved. The feature appeared with Grok 3 in early 2025 and behavior can be inconsistent across platforms.

Qwen 2.5 comes in sizes from 0.5B to 72B. The 7B model runs great on consumer hardware and punches above its weight. Excellent for coding tasks.ℹ️capability and hardware requirements vary substantially by size

Qwen 3 is Alibaba's latest open model family with hybrid thinking modes-switch between reasoning and fast response with `/think`. Sizes from 0.6B to 235B MoE, supports 119 languages. Fully open under Apache 2.0.ℹ️model sizes range from practical local options to very large MoE deployments

Qwen3-Coder-Next is Alibaba's latest coding model — MoE architecture with 80B total params but only 3B active, so it runs on consumer hardware (RTX 5090, 64GB Mac). 256K context, Apache 2.0 licensed, designed for agentic coding workflows. SWE-Bench Pro score of 44.3%.ℹ️tuned for coding rather than broad multimodal use

DeepSeek-V3.2 is a unified open model that combines chat and reasoning in one engine, replacing the separate V3 and R1 models. Features DeepSeek Sparse Attention for long-context, thinking-with-tools capability, and performance rivaling GPT-5. Distilled versions (1.5B-70B) run locally; full model needs enterprise hardware.ℹ️the full flagship models need serious hardware

LM Studio is a local desktop runtime product for discovering, loading, and chatting with open models through a GUI and optional local API.

Llama 3.3 is Meta's open-weight model. Completely free to run locally-you just need the hardware. The 70B model needs ~40GB VRAM for full precision, but quantized versions run on consumer GPUs.ℹ️higher-end variants still need significant VRAM

Llama 4 is Meta's first natively multimodal model-text and vision built-in. Scout has a 10 million token context window and fits on a single H100. Maverick rivals GPT-4o on benchmarks. Free for commercial use under 700M monthly users.ℹ️model size and context claims can exceed ordinary consumer hardware

Codestral is Mistral's code-specialized model at 22B parameters. Runs on a good consumer GPU (16GB+ VRAM recommended) and excels at code completion and generation.ℹ️consumer GPU use is realistic only with enough VRAM or quantization

Leanstral is Mistral's open-source code agent built for Lean 4 formal verification. Apache 2.0 weights with MCP tool-use training. Available self-hosted, via the Labs API, or in Mistral Vibe's agent mode.ℹ️specialized for Lean 4 formal verification, not general-purpose coding

Mistral offers several open models. Mistral Nemo (12B) runs on consumer hardware—great for local use. Mistral Large (123B) needs serious GPU power. Note: both are deprecated on hosted platforms (Vertex AI, La Plateforme) in favor of Mistral Small 4, but open weights remain available for self-hosted via Ollama and LM Studio.ℹ️runtime practicality depends heavily on which model size is chosen

Mistral Small 3 is a 24B model that matches 70B performance while being 3x faster. Fits on a single RTX 4090 or a MacBook with 32GB RAM. 81%+ on MMLU, 128K context in v3.1+. Fully open under Apache 2.0.ℹ️consumer use is realistic but still depends on memory and quantization choices

Mistral Small 4 is a 119B Mixture-of-Experts model with only about 6–6.5B active parameters per token. It combines instruct, reasoning, vision, and code in one open model. Apache 2.0 license, but requires datacenter GPUs (4× H100 minimum)—this is not a consumer-hardware model. Also available via the Mistral API as `mistral-small-2603`.ℹ️datacenter-class hardware required: minimum 4× H100 or 2× H200 or 1× DGX B200

Ollama is a local runtime product for downloading, serving, and interacting with open models through a CLI, desktop app, and local API.