ModelBytes

AI model release monitor for Telegram. Tracks new models from OpenRouter, Ollama, and Hugging Face, then posts daily curated digests to the @ModelBytes channel at 16:00 UTC.

Architecture (v3 — inline primary)

A small Python service on Railway. No claude.ai / Claude Code dependency — editorial taste is produced inline by an OpenAI-compatible writer model, grounded by cited web research:

• monitor.py — the publisher. A daily 16:00 UTC Railway cron that fetches OpenRouter / Ollama / HuggingFace, filters (is_noise_model / is_significant_release / is_stale_release), dedupes vs Postgres, collapses same-family variants, enriches from HF model cards, then has the writer model emit a format-v3 digest and posts it to Telegram + Slack.
• Inline writer — MODELBYTES_LLM_MODEL (production: deepseek-v4-pro on Ollama Cloud) with MODELBYTES_LLM_MODEL_FALLBACK (gpt-oss:120b). Grounded by Parallel.ai web research (MODELBYTES_PARALLEL_API_KEY) so it writes from cited sources, not training knowledge.
• Content gate — validate_digest_for_publish rejects anything that would harm the channel (stray <, unbalanced tags, floods, stale dates) before it reaches Telegram.
• pending/<TODAY>.txt — a write-back cache of what was published, read by tomorrow's cross-day fact-consistency check. (Earlier docs describe a claude.ai "curator" that wrote this file; that layer is retired — see docs/architecture.md § "How we got here".)

See docs/architecture.md for the full design and docs/operations.md for runbooks (rotating the bot token, manually triggering a post, reading publish_runs). The digest format (identity tiers + availability tags) is specified in docs/superpowers/specs/2026-06-10-builder-digest-format-v3-design.md. docs/vm-deployment.md and docs/structured-data.md cover a retired deployment path and the Postgres-first data roadmap.

Digest format (v3)

Identity tiers say what kind of model it is; a per-entry tag says how you can use it today:

• 🔓 OPEN FRONTIER / 🔒 CLOSED FRONTIER / 🎯 SPECIALIZED / 🏠 LOCAL / 👀 WATCH
• Every entry: bold name → italic differentiator sentence → hard facts → ⚡ API / 📦 weights availability tag → source link
• Lifecycle moves count: weights landing (WATCH graduations), big price changes, Ollama/OpenRouter arrivals
• Footer: Total: N items tracked today

Features

• 🔓/🔒 Open source vs proprietary classification
• ⭐ High-performance model detection
• ✨ Unique trait tagging (long_context, reasoning, multimodal, MoE)
• 📊 Benchmark scores when available
• 💸 Pricing info for API models
• 🗄️ PostgreSQL state persistence (required — set DATABASE_URL)
• 🚦 Duplicate-post protection via a posted_digests ledger
• 🔁 Retrying source fetches with a consistent ModelBytes user agent
• 🤖 Inline editorial digest via an OpenAI-compatible writer model (Ollama Cloud) grounded by Parallel.ai cited web research — no claude.ai / Anthropic dependency
• 🛡️ Reliability: every run is recorded in a publish_runs audit table, and failures or degradation alert the operator (Telegram DM, Slack fallback). The inline writer is the everyday path; pending/<TODAY>.txt is a write-back cache, not a deploy-timed handoff.

Deploy to Railway

(No public template — set up manually via the steps below.)

Manual Setup

1. Create Railway project
2. Add PostgreSQL (Railway provides DATABASE_URL)
3. Set environment variables:
   • TELEGRAM_BOT_TOKEN — From @BotFather
   • TELEGRAM_CHANNEL_ID — Your channel ID (use @getidsbot to find it)
4. Deploy

Local Development

# Clone
git clone https://github.com/SovereignSignal/modelbytes.git
cd modelbytes

# Setup
python3 -m venv venv
venv/bin/pip install -r requirements.txt

# Create .env
cp .env.example .env
# Edit .env with your tokens

# Run
python3 monitor.py

Running Tests

venv/bin/pip install -r requirements-dev.txt
venv/bin/python -m pytest tests/ -v

Environment Variables

Variable	Description	Required
TELEGRAM_BOT_TOKEN	Bot token from @BotFather	✅
TELEGRAM_CHANNEL_ID	Telegram channel ID	✅
DATABASE_URL	PostgreSQL connection string (Railway auto-sets; required for posting — --preview mode runs without it)	✅
MODELBYTES_LLM_KEY	API key for the deterministic-fallback LLM summarization step (OpenAI-compatible). Falls back to OPENAI_API_KEY then OPENROUTER_API_KEY if unset. Without any of these, the fallback path produces a template-only digest (no LLM editorial).	❌
MODELBYTES_LLM_MODEL	Model name for fallback summarization. Default: gpt-4o-mini.	❌
MODELBYTES_LLM_URL	API base URL. Default: https://api.openai.com/v1. Set to OpenRouter or another OpenAI-compatible endpoint to switch providers.	❌
MODELBYTES_HTTP_RETRIES	Source fetch attempts for transient failures. Default: 3.	❌
MODELBYTES_HTTP_BACKOFF_SECONDS	Base retry delay for source fetches. Default: 1.0.	❌
MODELBYTES_USER_AGENT	User-Agent sent to model source APIs. Default identifies ModelBytes.	❌
MODELBYTES_ADMIN_CHAT_ID	Telegram chat to DM on publish failures or degradation (operator alerts).	❌
MODELBYTES_OPS_SLACK_CHANNEL_ID	Slack channel used as a fallback for operator alerts when the Telegram DM can't be reached.	❌
MODELBYTES_HEARTBEAT_URL	Dead-man's-switch ping URL (e.g. healthchecks.io) — the only signal that catches "the cron never fired".	❌
MODELBYTES_PENDING_GRACE_SECONDS	How long the publisher waits for a late curator digest before falling back. Default: 600.	❌
MODELBYTES_ALLOW_SEED	Set to 1 to let the fallback path seed an empty models table (otherwise it refuses, to guard wiped/migrated state).	❌

These power the inline writer, which is the everyday digest path (the retired claude.ai curator layer did not use them). The writer has a primary + fallback model; if the primary (MODELBYTES_LLM_MODEL) returns empty, it tries MODELBYTES_LLM_MODEL_FALLBACK and alerts the operator that the primary was unavailable.

Sources

• OpenRouter — 400+ models with pricing
• Ollama — Local LLM models
• Hugging Face — Open weights and research models

See docs/source-growth.md for the source expansion rubric and candidate pipeline.

License

MIT