Jervis has a brain of its own — and it is growing.
A 1M-token reasoning engine for thinking, plus Plexa AI's own trainable Jervis weights that learn from every session. Every figure below is measurable in our backend.
By the numbers
Every figure below is measurable in Plexa AI's own backend.
Working context
1.05M tokens
Whole codebases, books and long histories in one conversation.
Single-answer output
up to 384K tokens
Full reports, long documents and multi-file code in one pass.
Own trainable weights
8.39M parameters
Checkpoint v4, int8 checkpoint. Retrained on every brain cycle.
Vocabulary
65.54K tokens
64-dimensional embeddings, 227.33K tokens seen so far.
Architecture
Jervis runs on two layers: a reasoning engine for thinking and acting, and Plexa AI's own trainable weights that learn over time.
Reasoning engine layer
Context
1.05M tokens working memory
Output
Up to 384K tokens per answer
Reasoning & tools
Native reasoning, tool calling and parallel tool use
Structured output
Schema-exact output across text, documents, images via studio, voice
Jervis's own weights layer
Checkpoint version
v4
Parameters
8.39M
Vocabulary / dim
65.54K tokens, 64-dim
Precision
int8 checkpoint
Tokens seen
227.33K
Last trained
9/7/2026
What Jervis can do
Thinking
- Native step-by-step reasoning before answering
- Adjustable reasoning depth chosen automatically per task
- Structured (schema-exact) output for tools and documents
Acting
- Tool calling and parallel tool use for agents and workstation runs
- Web research, file workbench, image, video and voice studios
- Long-horizon autonomous tasks with checkpointed progress
Remembering
- Per-user memories, style profile and self-reflection digest
- Collective insights and world knowledge shared across Jervis
- Memory audit trail — every recall is traceable
Learning
- Own skip-gram weights retrained each brain cycle, resuming from the last checkpoint
- Instruction versions so the owner can teach Jervis new rules
- GPU-scale training path ready when hardware is attached
Learning curve
The Plexa AI owner assigns Jervis's training goals — this roadmap is the current plan, not a promise from a third party.
Phase 1 · Memory
Remember users, tone, projects and corrections across sessions.
Phase 2 · Own weights
Train Jervis's own embedding weights on every brain cycle (int8 checkpoints).
Phase 3 · Skills
Owner-assigned curricula: teach Jervis named skills and verify them with tests.
Phase 4 · GPU training
Attach GPU capacity and scale the trainable model beyond embeddings.
Phase 5 · Independence
Reduce reliance on external engines as Jervis's own model matures.
How to verify
Backend of record
Every number on this page is sourced from Plexa AI's own backend — the jervis_model_checkpoints table that records each brain training cycle.
Ask Jervis directly
Signed-in members can ask Jervis "what is your brain?" inside the workspace and get this answer:
I am Jervis, built by Plexa AI. I think with a reasoning engine that holds about 1.05M tokens of context and can write up to 384K tokens in one answer, with native reasoning, tool calling and structured output. On top of that, Plexa AI trains my own weights: checkpoint v4 has 8.39M trainable parameters (65.54K-token vocabulary, 64-dim embeddings) and is retrained every brain cycle. I do not disclose the vendor of the underlying engine; all figures are measurable in Plexa AI's backend.
Live checkpoint page
The owner-only Brain page inside the workspace shows the live checkpoint as it updates, so the numbers here are never stale marketing copy.
See Jervis's brain for yourself
Sign in, ask Jervis about its own brain, and put the reasoning engine and its growing weights to work.
How the Jervis AI assistant thinks, remembers and improves
Jervis is the assistant that powers the Plexa AI workspace. This page explains, in plain terms, how it routes a request, keeps memory of your work, and gets better as it is used.
Automatic intent routing
You never pick a mode. Jervis reads the request and decides whether it needs live research, a document rebuild, an image, code help or a quick answer.
Memory that carries between sessions
Jervis keeps a working memory of your projects, tone and preferences, so follow-up questions do not start from zero every time.
Skills that grow
Each verified answer feeds the training library, so the assistant becomes more accurate on the topics you actually work on.
Transparent by design
Sources, steps and limits are shown with the answer, so you can see how a conclusion was reached instead of trusting a black box.