Releases: Somi-Project/Somi
V0.92 Mineralization
SOMI v0.92 - Mineralization
Mineralization is the release where SOMI stops feeling like a promising framework and starts feeling like a real local AI operating system.
This version is about hardening, densifying, and refining the core. More capability. Better structure. Stronger control. Cleaner operator experience.
Under pressure, SOMI did not crack. It crystallized.
Release Theme
v0.92 Mineralization is a foundation release.
It takes SOMI's major systems and turns them into something more durable:
- more modular
- more secure
- more observable
- more self-hosted
- more usable for real people
This is the chapter where the framework becomes far more credible as a daily driver for local AI.
Highlights
SOMI Now Feels Like A Real AI Workstation
SOMI now delivers a much more complete local operator experience with:
- Control Room
- Coding Studio
- Research Studio
- Speech controls
- Node Manager
This is no longer just a command loop with extra tools attached. It is a structured desktop-first environment for running your own AI system.
Coding Mode Became Serious
SOMI now includes a much stronger local coding stack with:
- managed coding workspaces
- guarded file operations
- Python-first coding workflows
- multi-language workspace support
- verify loops and scorecards
- a dedicated coding cockpit
The goal was simple: make local coding in SOMI feel closer to a serious coding agent instead of a novelty feature.
Research Got Much Stronger
Research mode now goes beyond plain search-and-summarize.
SOMI now supports:
- research supermode
- evidence graphs
- grounded exports
- contradiction-aware synthesis
- longer-lived research workflows
This release makes SOMI much more competitive as a self-hosted research system.
Speech Was Rebuilt
The speech layer was overhauled into a cleaner, more practical local-first pipeline.
SOMI now uses:
- local TTS via
pyttsx3 - local STT via
faster-whisper - diagnostics and health checks
- desktop speech controls
The result is a speech system that is easier to understand, easier to install, and much more believable for everyday use.
Security And Control Matured
Mineralization pushed hard on trust and operational safety.
Major gains include:
- approval-aware execution
- scoped remote behavior
- gateway and node mesh foundations
- runtime audit trails
- release gate and framework freeze tooling
- stronger control-plane visibility
SOMI is built to be powerful without pretending risk does not exist.
The Backbone Is Stronger
A huge amount of work in this release went into the parts most users never see directly, but feel immediately:
- split monoliths
- cleaner package boundaries
- workflow runtime
- subagent runtime
- ontology-backed actions
- persistent state plane
- memory search and compaction resilience
- tool registry and execution backends
- skill forge and marketplace foundations
This is the kind of work that makes a framework survive growth.
What Makes v0.92 Different
Earlier versions of SOMI proved the idea.
v0.92 Mineralization proves the shape of the product.
This release is where SOMI becomes:
- a stronger self-hosted desktop AI
- a stronger local coding partner
- a stronger research framework
- a stronger automation and control system
- a stronger foundation for future packaging and onboarding
Pillars Reaffirmed
Mineralization keeps SOMI aligned with its core design pillars:
- User experience
- Security
- Flawless functions on consumer-grade hardware
- Fast
- Modular
This release was shaped around those values, not around feature count for its own sake.
User-Facing Wins
For everyday users, this release means:
- a more futuristic but practical desktop interface
- a more trustworthy local experience
- stronger speech, coding, research, and OCR capabilities
- more obvious control over what the system is doing
- less dependence on cloud-style patterns
For builders and power users, this release means:
- better modularity
- clearer extension points
- better execution boundaries
- richer observability
- a far more serious platform to build on
Under The Hood
Key systems strengthened in this release include:
- PySide6 desktop migration and cockpit polish
- coding toolbox and workspace runtime
- speech pipeline redesign
- workflow runtime and subagents
- ontology and action model
- gateway, pairing, and node mesh direction
- release gate, freeze, replay, and evaluation harnesses
- memory and session-search improvements
This was not a cosmetic update. It was structural.
Known Reality
SOMI v0.92 is a major step forward, but it is still a living framework.
The next major focus after Mineralization is:
- A secret for now as we take a break
The core is now much stronger. The next chapter is making that power easier for more people to access.
Closing
SOMI v0.92 - Mineralization is about turning scattered capability into dense capability.
More than adding features, this release makes the system feel more coherent, more sovereign, and more ready.
If previous SOMI builds felt like excavation, this one feels like formation.
SOMI is becoming the kind of local AI framework people can actually live in.
v0.91 — The Executive Brain Release
Somi has crossed a structural threshold.
With this upgrade complete, the system now resembles an executive cortex — deliberate, persistent, and selectively proactive. This release is not about features. It is about cognition.
What Changed?
Structured Cognition (Prefrontal Layer)
All meaningful outputs are now deterministic artifacts.
No loose reasoning. No unpredictable drift.
Persistent Memory (Hippocampal Layer)
Tasks link to projects. Projects link to goals.
Somi understands trajectory, not just status.
Guarded Execution (Basal Ganglia Gate)
Capabilities exist — but nothing executes without approval.
Every action is scoped, logged, and revocable.
Strategic Intelligence (Control Network)
Tradeoff analysis, plan revision, risk scoring — structured, auditable, non-autonomous.
Selective Attention (Salience Network)
Proactivity is now controlled by a Signal Strength Index:
Notify
Include in brief
Log only
Interrupt budgets, escalation ladders, quiet hours, and fingerprint suppression prevent spam.
Morning and evening briefs are:
Action-focused
Threshold-driven
Fully preference-controlled
Reversible via natural language
“Forget weather today.”
“Only alerts.”
“Move my morning brief to 9.”
All handled structurally
Where We Stand:
- Structured cognition
- Persistent state
- Controlled execution
- Personalization
- Life awareness
- Strategic intelligence
- Attention discipline
Somi is now a disciplined, local-first AI operating layer running on 3–9B models.
Structured intelligence.
v0.9 — Local Sovereignty Edition
Major Improvements:
- Refined agent routing
- Improved RAG injection
- Social automation stability improvements
- Memory reliability upgrades
Somi is now closer to a true local-first agent OS.
Next focus:
- Deterministic memory system
- Skill registry
- Execution sandbox