Roadmap

Roadmap

Shipped

Satellite	Package	Status	Modality	Description
transcriber	rvrb-transcriber	alpha	AUDIO→TEXT	AI-powered audio transcription
verify	rvrb-verify	alpha	TEXT→TEXT	Claim verification engine (two-phase pipeline)
transform	rvrb-transform	alpha	TEXT→TEXT	General-purpose text transformation

Planned — Tier 1 (Core Modalities)

#	Satellite	Modality	Model	What
1	see	IMAGE→TEXT	qwen3.7-plus	Image understanding, OCR, visual description
2	hear	AUDIO→TEXT	qwen3.5-omni-plus	Multimodal audio comprehension
3	speak	TEXT→AUDIO	cosyvoice-v3.5-plus	Text-to-speech via DashScope API

Planned — Tier 2 (Composable Pipelines)

#	Satellite	Modality	Model	What
4	summarize	TEXT→TEXT	qwen3.6-flash	Long-document summarization (1M context)
5	translate	TEXT→TEXT	qwen-mt-plus	High-quality translation
6	extract	TEXT→TEXT	qwen3-coder-plus	Structured data extraction
7	reason	TEXT→TEXT	qwq-plus	Multi-hop reasoning and chain-of-thought

Planned — Tier 3 (Specialized)

#	Satellite	Modality	Model	What
8	codegen	TEXT→TEXT	qwen3-coder-plus	Code generation with tool calling
9	review	TEXT→TEXT	qwen3.7-plus	Code review against standards
10	detect	IMAGE→TEXT	qwen3.7-plus	Object detection, scene analysis
11	subtitles	AUDIO→TEXT	qwen3-asr-flash	Subtitle generation (SRT/VTT)

Infrastructure

Item	Status	Description
Satellite protocol v2	done	Multimodal I/O contracts, provider resolution, mock patterns
Scaffold script	done	scripts/scaffold-satellite.py — generates v2 skeletons
Qwen fallback rotation	done	scripts/qwen_fallback.py — 81-model quota rotation
opencode Qwen config	done	Multimodal models in /models, auto-fallback

Model Strategy

The Qwen free quota (103M tokens across 81 LLM models, expiring Sept 2026) is the fuel. Each satellite targets a specific model family.

Tier	Models Used	Token Budget
Tier 1 (core)	qwen3-coder-plus, qwen3.7-plus, qwen3.5-omni-plus	~4M tokens
Tier 2 (compose)	qwen3.6-flash, qwen-mt-plus, qwq-plus	~4M tokens
Tier 3 (specialize)	qwen3-asr-flash, qwen3.7-plus	~2M tokens
Development (opencode)	qwen3-coder-plus	~20M tokens

See Model Catalog for the complete list of available models.

North Star

Every satellite is a composable, MCP-native Python package. No monolith. Each exploits a specific Qwen model family capability. Together they form a multimodal toolkit for text, audio, image, video, and speech — all running on zero-cost Alibaba Cloud free quota.