One reasoning core. Every place the work happens.
Every Vexoo product runs on the same hallucination-resistant core — small models that reason over your own data, cite every answer, and never leave your perimeter.
Four engines under everything.
Every Vexoo product is built on the same stack — the reasoning, the memory, the tabular analyst, and a remarkably small engine.
THE BRAIN
SRA · MARLIN
Structured Reasoning Adapters
Fine-tuning teaches a model what to say. Our adapters teach it how to think, injecting step-by-step reasoning into the architecture itself. That’s why a 1–3B model reasons like one many times its size.
THE MEMORY
Knowledge Pyramid
Hierarchical retrieval, not flat chunks
Flat RAG shreds your documents into interchangeable chunks and hopes the nearest match is the right one. The Pyramid keeps structure: granular facts at the base, concepts and cross-references in the middle, synthesised themes at the top — every layer carrying provenance.
THE ANALYST
Quantix
Tabular reasoning, row and column aware
Most retrieval systems flatten a table into a wall of text and lose the thing that made it a table. Quantix reasons over structure — financial statements, spec sheets, comparison matrices — reading down a column and across a row the way an analyst does.
THE ENGINE
Apex 3B
A small model that punches up
Three billion parameters scoring 77.3 on GPQA Diamond, ahead of GPT-4.1 and GPT-4o on scientific reasoning, at roughly 40% the compute, on a single GPU, fully air-gappable.
Quantix, in practice
A table stays a table.
Flat-chunk retrieval smears a financial table into prose and loses the column headers. Quantix keeps the structure and reads down the column — then cites the exact cell it read.
Playground
LiveGrounded Enterprise Intelligence
Ask your own documents a question. Get a clear answer with the exact source.
The flagship platform. It ingests the documents, policies and data you already own and turns them into a queryable knowledge base where every answer is traced to the exact clause, page and paragraph it came from. Yard, its secure DocChat surface, puts the same engine in front of everyday users.
Built for: Compliance, legal, risk, policy and operations teams.
Capabilities
- Knowledge Pyramid, hierarchical retrieval with full provenance
- Evidence-first answers, every claim cited to its source
- Yard (DocChat), secure Q&A over private document corpora
- Multi-document reconciliation: statute beats policy, latest beats earlier
- Deploys on-premise, in your VPC, or fully air-gapped
- 22 Indian + 40 global languages, including code-mix
Quantix
LiveTabular Reasoning Engine
Ask questions of your spreadsheets and filings, not just your paragraphs.
The structured-data engine. Most retrieval systems flatten a table into a wall of text and lose the thing that made it a table. Quantix preserves rows, columns, headers and footnotes as structure, then reasons over them the way an analyst does — down a column, across a row, cited back to the exact cell.
Built for: Finance, research, procurement and analytics teams.
Capabilities
- Structure survives ingestion, rows, columns, headers and merged cells
- Reads down and across, year-on-year and cross-row comparison
- Unit and currency reconciliation across inconsistent sources
- Cited to the cell, every figure traces to table, row and column
- Reconciles a figure against the clause that governs it
Vexoo SRA
LiveStructured Reasoning Adapter · Apex 3B
Frontier-class reasoning on one GPU, inside your own network.
The Structured Reasoning Adapter runtime. MARLIN adapters inject step-by-step reasoning into the architecture itself rather than fine-tuning what a model says, which is why a 1–3B model reasons like one many times its size. Pairs with retrieval grounding, and ships OEM or on-prem.
Built for: Platform teams, OEMs and anyone deploying AI inside a perimeter.
Capabilities
- MARLIN adapter architecture, reasoning built into the model
- 77.3 on GPQA Diamond, at roughly 40% the compute
- Runs on a single GPU, fully air-gappable
- Hot-swappable adapter sets per domain
- OEM and on-prem licensing
Forge
LiveEnterprise AI Coding Assistant
An AI coding assistant your security team will actually approve.
A private-by-default coding assistant in two form factors — Forge Desktop and Forge TUI — sharing one engine, one account system and one policy layer. Public AI coding tools route your source through a third-party cloud and lock you to a single vendor's model. Forge does neither.
Built for: Engineering orgs under IP, residency or regulatory constraints.
Capabilities
- Private by default, code and prompts stay on your infrastructure
- Model-flexible, cloud, private cloud or fully on-prem, bring your own
- Two form factors, Forge Desktop and Forge TUI, one engine
- Cost control, route cheap tasks to cheap models automatically
- Enterprise-ready, SSO/SAML, RBAC and full audit trails
- Customised agentic builds for domain workflows, e.g. Forge · Sulzer
Built on the core.
Surfaces that take the same grounded engine somewhere specific — the open web, and the meeting room.
Vexoo.ai
Research-Grade Answer Engine
Search that shows its work.
A research-grade answer engine that routes every question through verified sources and returns source-anchored answers. The consumer face of the same grounded reasoning stack that powers our enterprise platform.
- Web Search & News Search, routed through trusted sources
- Scholar, simple and multi-hop academic reasoning
- Deep Research, long-form investigation with a full source trail
- Job Lens, role and market intelligence
- The same reasoning core as Playground, at consumer scale
Echo
AI Meeting Assistant
Every decision, traceable to the minute it was made.
Transcription, structured summaries and action items for meetings, plus Q&A over the recording where every answer points back to the timestamp it came from. The same grounding discipline as the rest of the stack, applied to what your teams actually said.
- Accurate transcripts with speaker attribution
- Structured summaries and extracted action items
- Meeting Q&A grounded to the exact timestamp
- Searchable across your full meeting history
The models underneath, licensable on their own.
If you want the reasoning layer without the platform — OEM, embedded or inside your own product — these ship independently.
Apex 3B
Reasoning Small Language Model
77.3 on GPQA Diamond, on a single GPU.
The model the SRA runtime is built around, licensable on its own. Three billion parameters trained and adapted for reasoning rather than recall — it scores ahead of GPT-4.1 and GPT-4o on hard scientific reasoning at roughly 40% the compute, and it runs entirely inside your perimeter.
- 77.3 GPQA Diamond, ahead of GPT-4.1 and GPT-4o on scientific reasoning
- 72.3 GSM8K multi-step maths · 59.8 MMLU
- Roughly 40% the compute of comparable systems
- Single GPU, fully air-gappable, no external calls
- Methodology shared on request
Nymble 350M
Edge Small Language Model
Intelligence that runs where the network doesn’t.
An ultra-small language model for on-device and embedded use. Built for constrained hardware where a round trip to a data centre isn’t an option — predictive maintenance inside a moving vehicle, industrial edge, field equipment.
- ~50ms on-device inference
- Runs in 6–8 GB RAM, validated down to a standard i5 laptop
- Zero connectivity required
- Automotive predictive maintenance and energy management
Vertical solutions.
Where a domain needs the whole answer rather than a platform, we ship it as its own solution on the same core.
Not sure which product fits?
Tell us the decision you can’t afford to get wrong. We’ll map it to the right deployment, and prove it on your own documents.