• Preview attachments: now you can click to attachments/images to preview them directly in the app, including text, code, CSV, Excel, Word, PDF, and other popular file formats.

• Haiku 5.5 — fastest, cheapest Claude 5.5. Best for high-volume work: summaries, classification, routing, subagents, live support.

• Mistral Large 4 — Mistral’s new flagship (aka Le Chonk). Strong coding, agents, and images; open weights coming soon.

• Nano Banana 2.1 — Google’s faster, cheaper image model. Better text, character consistency, and panoramas.

Your AI can remember context about you, your preferences, and the way you work - so you don’t have to repeat yourself in every new chat.

For users

Attachment/Generated files are now managed under About this chat → Files.

GPT Image 2.5 can now edit any image available in the current chat, instead of being limited to images from your latest message. The AI will automatically select the relevant images to use.

Interactive Canvas and Web App Builder can now use images directly from the chat. When you share an artifact, the images displayed inside it and are included as part of the shared artifact.

For plugin developers

Plugins can now receive files directly from the chat using the new typingmind-attachment parameter type together with the read_chat_attachments permission. Learn more

Image cards now support a new image.filename field. Base64 images returned by cards are also automatically saved as chat files. Learn more

Chat files can now be referenced directly in plugin output:

• HTML: tm-attachment://<id>

• Markdown: #tm-attachment:<id>

Learn more

For GitHub-hosted plugins, JavaScript functions can now be stored in separate files using the new codeFile field. Learn more

Behavior changes

previousRunOutput no longer includes output from respond_to_ai. Learn more

When importing a plugin from a private GitHub repository, TypingMind now uses the main branch by default. To import from another branch, add **`

• Added user group filtering to org analytics, allowing users to view tokens and messages broken down by user group, plus a new "User Groups" option when exploring analytics data.

• Integration API key generation is now restricted to account owners and admins; other roles can still view and revoke existing keys.

UXUI Improvements:

• Open "About this chat" in a convenient side panel, better view artifacts on both desktop and mobile.

• Improved the button UI in rendering results for artifacts, HTML, and code.

• Improved support for reading PowerPoint (.pptx) attachments.

• GPT-6 Sol — OpenAI’s mid-tier GPT-6 (below Astra). Complex coding and agents, much of Astra’s strength at half GPT-5.6 Sol’s price. $2 / $10 per M tokens.

• GPT-6 Luna — Cheapest/fastest GPT-6. High-volume focused tasks (extract, classify, summarize, simple code). $0.10 / $0.50 per M tokens.

MiMo-V2.6-Pro — Xiaomi open-weight flagship (MIT). 1.02T MoE / 42B active, 1M context, text+image+video+audio. Top-tier open model. API ~$0.44 / $0.87 per M tokens.

MiMo-V2.6-Flash — Cheaper Xiaomi twin. 309B / 15B active, same 1M context and modalities. Close to Pro on many agent tasks. API $0.14 / $0.28 per M tokens.

You can now save your favorite plugins, skills, and MCP servers into named presets and switch between them with a single click.

Create presets for different workflows like Frequently Used, Coding, Support, Sales, or Research — so you always have the right tools enabled for the task at hand.

Also in this update, you can search plugins, disable all plugins instantly, and export/import presets across devices.

• Side Chat Model Selection: Users can now choose which model powers the side chat "Ask..." feature (highlight text in a chat → select Ask...), either via a new setting or directly from the side chat popup.

• Temporary Chat for Agents: Added support for temporary chat when using agents.

• New filters and sorting on the Users page, including Last Login

• More bulk actions for managing users and groups

• Smarter exports based on your selection or active filters

• Search and sorting for Groups

• Safer “Select all” behavior when adding group members

• Smoother import previews for large user lists

We’ve also improved performance and scalability for organizations with large numbers of users.

• Support stopping and regenerating messages in multi-model responses

• Support pasting files and large text as file attachments when replying in multi-model threads.

• Show statistics such as token usage and estimated cost for each model.

• Fork a chat directly from a multi-model response message.

• When finalizing a multi-model response, you can now click Edit finalize instruction to customize the prompt used for that specific finalize request before picking the model to generate it.

Note: The default finalize instruction can now be customized in Settings → Advanced Settings → Internal Prompts under a new "Finalize multi-model responses" section (with a "Reset to default" option), synced across your devices via cloud sync.

• Claude Fable 5.1: Anthropic’s most advanced model for demanding reasoning, long-horizon coding, deep research, and complex knowledge work.

• Gemini 3.8 Flash: Google’s newest high-speed workhorse, built for powerful reasoning, coding, autonomous agents, and complex workflows—while keeping Flash-level speed and efficiency.

• Muse Spark 1.3: Meta’s latest Muse model, improved for agentic workflows and coding, with a stronger ability to handle long-running tasks and multiple workflows in one conversation.

Try them now on TypingMind!

We also added Qwen3.8 Flash - latest model from Alibaba:

Bug Fixes and Improvements:

• Added native PDF capability detection with a new "Support PDF Input" toggle when manually importing custom models.

• Fixed an issue where synced images could not be used with the DeepSeek provider.

• Fixed an issue where attachments were not accessible in forked chats.

• Add skills:

• Enable skills / control skill visibility to specific user groups:

• Assign skills to Agents:

• Ox Alpha (Free) is available via OpenRouter and OpenCode Go

Bug fixes and improvements:

• Fixed an issue where new multi-model chats would revert to a single model after switching chats.

• Added support for the Responses API type for OpenCode Go.

• Added support for external image URLs when using OpenAI Responses models.

• Update DeepSeek V4 Flash 0731 - the latest version across providers:

App improvements:

• Shared chats now support displaying multiple responses.

• Your device will now stay awake while a response is being generated, preventing long responses from failing due to the screen turning off.

Bug Fixes and Improvements:

Bug fixes:

• Fixed an issue where plugins couldn't properly access synced blobs.

• Fixed an issue where images attached to shared chat links appeared broken for people viewing the link without an account.

• Fixed an issue where generated images would disappear after the streaming response finished.

Improvements:

• Added support for using external PDF URLs with OpenRouter models, instead of requiring inline file data.

• Removed the redundant "TypingMind AI Gateway & Org ID" entry from Settings → Advanced Settings, since these settings have already moved to the Models tab.

• Uninstalling a plugin now also clears its saved settings and enabled/disabled state, so reinstalling it starts fresh instead of restoring stale values.

• GPT-5.6 Sol — the flagship model, designed for complex reasoning, advanced coding, long-running agent workflows, scientific research, and cybersecurity tasks.

• GPT-5.6 Terra — a balanced, lower-cost model for everyday professional work, offering strong capability at approximately half Sol’s price.

• GPT-5.6 Luna — the fastest and most cost-efficient option, ideal for high-volume requests, quick answers, lightweight coding, summarization, and routine automation.

Enable and use it on TypingMind now!

Bug Fixes and Improvements:

• Issue with Temporary Chat still being saved within folders: now Temporary Chat can no longer be turned on when the current chat belongs to a project/chat folder.

• Replaced DALL-E plugin access with GPT Image Editor for all plans.

• Claude Sonnet 5: Anthropic’s most agentic Sonnet model yet.

• Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image): Google’s fastest and most cost-efficient image model.

New models have also been updated across providers such as OpenRouter and Vercel AI Gateway.

In this update, we also added a temporary fix for the CORS issue affecting the Perplexity Search plugin and Perplexity models.

The Fugu Ultra model is said to match the performance of Fable and Mythos.

Try now!

• Disable drag-and-drop uploads on Admin Panel when Upload Documents option is disabled. Now, uploaded documents are also counted toward usage limits.

• Fixed image uploads for AWS Bedrock Claude models by properly resolving URL-based images and showing clearer re-upload errors.

• Fixed PDF/image issues for Fireworks and Novita by disabling unsupported PDF support.

Improved

• Refreshed model catalogs and metadata across Chutes, DeepInfra, Fireworks, Mistral, Moonshot, Novita, OpenRouter, and Vercel.

• Improved OpenRouter API key verification and model search/import messaging.

• Improved MCP server display and connection reliability, including device-specific server filtering, better initialization, SSE timeout handling, and OAuth URL generation.

Added / Deprecated

• Added/promoted new models including GLM 5.2, Kimi K2.7 Code, MiniMax-M3, Qwen 3.7 Plus, and Fireworks fast-router variants and more.

• Deprecated older/free models.

• Added automatic model capability checks before saving a custom model. TypingMind can now test model connectivity and detect support for features such as system role, streaming, tools, vision, thinking, prompt caching, supported parameters, and provider tools.

• Added support for storing cache read/write pricing fields for custom models.

You can now connect your API keys and use models from these providers directly in TypingMind.

Go to Models → API Keys to set up your preferred provider and start using the available models.

TypingMind now supports Agent Skills — bundled procedural knowledge that helps AI perform specific tasks more effectively, such as SEO audits, copywriting, code reviews, brainstorming, and more.

Each Skill is packaged around a SKILL.md file, which contains task-specific instructions and optional reference files, following the open Agent Skills standard.

How to install Skills on TypingMind:

1. Go to Plugins.

2. Under the Store section, select Skills.

3. Install available Skills from the Store, or import your own Skill using a GitHub URL or ZIP file.

How to use Skills on TypingMind:

Enable the Skills you want to use, then start chatting. TypingMind will use the selected Skills to provide more specialized and task-focused responses.

Learn more about Skills: https://docs.typingmind.com/skills

Note: Skills are now available on TypingMind.com (Personal version). Support for TypingMind Teams coming soon.

✨ Also in this update:

• Added Claude Opus 4.8 and Opus 4.8 (Fast) via OpenRouter.

• Updated model provider pricing/metadata for several catalogs, including DeepSeek, Deepinfra, Groq, Mistral, and OpenRouter models.

• Added new Deepinfra models and OpenRouter models.

• Deprecated/deleted models are now kept as deprecated so old chats can show model status more clearly.

Improvements

• Updated OpenRouter and Deepinfra model metadata to reflect latest limits, pricing, and supported capabilities, including context/output limits, pricing, PDF support, and prompt caching support.

• Updated OpenRouter Anthropic latest aliases to use Claude-specific settings for better reasoning effort and prompt caching support.

• Expanded OpenRouter reasoning effort values: none, low, medium, high, xhigh.

• Removed deprecated DALL-E 3 builtin plugin as this image model is deprecated by OpenAI. Please switch to GPT Image Editor for image generation.

Fixed

• Fixed Fireworks provider issue where previous user input or stale/duplicated assistant content could be echoed when using the OpenAI Responses API format.

This models is optimized for coding proficiency and parallel agentic execution loops.

Support all provider settings for profiles

• You can now set up API keys for different profiles from all available providers

• Older profile API key assignments are automatically migrated to the new provider settings format.

⚙️ Bug fixes and improvements

• In multi-model chats, clicking a model header/tab now scrolls directly to that model’s latest assistant message. This makes it easier to continue reading the newest response without scrolling through earlier replies.

• Fixed an issue where switching model tabs on mobile could fail to scroll to the latest message if the response card had not mounted yet.

• Removed unsupported metadata fields from tool calls before sending requests to OpenAI-compatible providers.

• Fixed provider API errors caused by extra tool call fields such as output_item_id, call_id, and annotations.

✨ New

Redesigned Admin Analytics

• The analytics dashboard now has two tabs: Overview for high-level insights and Exploration for deeper analysis.

• New metrics have been added, including Characters, plus breakdowns by users, models, and agents, with improved charts, styling, and tooltips.

• CSV export for analytics data has been centralized for a more consistent experience.

• Date ranges are now inclusive, normalized to UTC day boundaries, and consistently limited to 1 year.

Exploration mode now supports filtering by metric, including user-level filtering where permitted.

🛠 Improvements

Web App Builder & Interactive Canvas:

• AI can now update existing HTML using small patch-style edits instead of regenerating the full document, making follow-up edits faster and more token-efficient.

Artifacts side panel — version labels:

• When a plugin has multiple tool functions, artifact versions are now correctly grouped into one continuous stack and no longer display "undefined - Version N".

⚙️ Bug Fixes

• Fixed DeepSeek API handling so reasoning_content is preserved for supported models.

• Fixed an issue where MCPs/tools did not work with some providers, including Mistral.

• Improved cross-provider tool/function calling compatibility by removing unsupported annotations from function specs before sending tool definitions.

• Improved the Organization dashboard Recent Activity feed with clearer audit log formatting, names, and layout.

Here's how it works

• Go to Settings → Cloud Sync & Backup → Refer friends

• Copy your personal referral link (typingmind.com/r/<code>). For example:

  https://www.typingmind.com/r/S4G6TB5G

• Share it with friends via Email, X, or LinkedIn — or anywhere else

When they purchase a license through your link, you both get 0.5 GB added to your accounts instantly.

Notes:

• Available for individual TypingMind accounts (typingmind.com)

• No expiry on earned storage — keep it for life

• Your friend must use your link before purchasing for both of you to receive the reward

You can now easily add models from:

• OpenRouter

• DeepInfra

• DeepSeek

• Groq

• Mistral

• Moonshot

• Perplexity

• xAI

• Z.ai

More providers and models will be added soon.

We also moved API keys setting from Settings → API Key to Models → API Keys.

To set up API keys, go to Models from the left-side panel, then switch to the API Keys tab and enter your API key.

You can also go to Models and click the gear icon next to each AI provider to update its API key.

Also in this update:

• Added support for custom endpoints for OpenAI Responses API requests.