v1.5.0

What's New

Multi-Image Context for Vision Models

• Previous images (up to 5) are now sent alongside the current image when memory is enabled
• Images are sent in chronological order (oldest first) so Molmo can correctly reference them
• Prompt includes ordering annotation so the model knows which image is "previous" vs "current"
• Image history is cleaned up when memory is disabled or /clear is used

This fixes the issue where Molmo would confuse the order of images when asked to compare them.

v1.4.0

What's New

Web2API Access Token Support

• New OLMO_WEB2API_TOKEN env var for authenticated requests to web2api
• Required since web2api v0.6.0 protects all routes by default
• Bearer token sent as Authorization header on all requests

Docs

• Updated README with /search command documentation

v1.3.0

What's New

Telegram formatting

• OLMo's markdown responses now render as proper Telegram HTML (bold, italic, code, links, headings)
• Falls back to plain text if HTML parsing fails
• Suppresses link previews for cleaner output

Response cleanup

• Truncates hallucinated multi-turn artifacts (OLMo sometimes generates fake follow-up conversations)
• Strips raw <function_calls> and role markers from responses

v1.2.0

What's New

/search command

• Web search is now opt-in via /search <query> instead of auto-triggering on every message
• Automatically falls back to a tool-capable model (olmo-32b) when the current model doesn't support tools

Persistent typing indicator

• "Typing..." now stays visible the entire time the model is generating a response (re-sent every 4 seconds)

Tool model awareness

• New TOOL_MODELS config — only olmo-32b and olmo-7b support Allen AI's native tool calling
• Removed inline mode (bad fit for LLM queries)
• Added web_reader tool support

v1.1.0 — Point Overlay & Web Search

What's New

🎯 Point Overlay (Molmo 2 Vision)

Send a photo with a caption like "Point to the eyes" and Molmo 2 will return the image with colored markers drawn on the detected points.

• Smooth anti-aliased markers (4× supersampled, LANCZOS downscaled)
• Numbered colored dots (red, blue, green, orange, purple, pink, cyan, yellow)
• White border ring with soft glow for visibility on any background
• Auto-scaled marker size based on image dimensions
• Caption shows point labels; redundant text messages suppressed

Prompts that trigger pointing:

• "Point to the cat"
• "Find the eyes"
• "Where is the nose?"
• "Show me where the people are"

🔍 Web Search (All Text Models)

All text models now have access to Brave Search via the MCP tool bridge. The model can autonomously decide to search the web when it needs current information.

• Passed via tools_url parameter on every text model request
• Vision models (Molmo 2) excluded — they don't need web search
• Configurable via OLMO_TOOLS_URL environment variable

Dependencies

• Added Pillow>=10.0 for image overlay rendering

Full Changelog: v1.0.0...v1.1.0

v1.0.0 — OLMo Telegram Bot

Features

🤖 Multi-Model Chat

• OLMo 3.1 32B Instruct — default model
• OLMo 32B Think — reasoning model
• OLMo 7B — lightweight model
• Tülu 8B / 70B — instruction-tuned alternatives
• Switch models anytime with /olmo32b, /think, /olmo7b, /tulu8b, /tulu70b

👁 Vision (Molmo 2)

• Molmo 2 8B — image and video understanding
• Molmo 2 8B Tracking — 8fps video tracking
• Send a photo or video with a caption to analyze it
• Auto-switches to Molmo 2 when media is sent on a text-only model
• /molmo2 and /molmo2track to switch explicitly

🔍 Web Search

• All text models have access to Brave Search via MCP tool bridge
• Models can decide to search the web when they need current information
• Powered by Web2API MCP integration

💬 Other

• Inline mode — @your_bot query in any chat
• Conversation memory — optional per-user history (/memory)
• Access control — restrict to specific Telegram user IDs
• Configurable via environment variables

Setup

See README for installation and configuration.