A research project in local artificial intelligence
Frozerti
Researching AI that can do more within the resources it has.
We study how to preserve useful language-model abilities when memory and computation are limited. We develop experimental solutions, test them on new tasks and keep a public journal of observations, corrections and rechecks.
Stage: research and experimental development
The problem
Useful AI within a device’s resources
Memory, computation and response time limit what a local assistant can do. The device must also support the conversation context and other software.
We look for ways to use fewer resources while keeping answer quality. Every solution is tested in comparable trials.
What we account for
- 01
Memory
Available memory must support the model, its context and the software running alongside it.
- 02
Data movement
Moving data also takes time and can become a limiting factor.
- 03
Response time
An assistant must solve tasks with sufficient quality in a reasonable time.
The Frozerti idea
More available abilities. Less in GPU memory.
We study a model where only a small portion of its experts needs to reside in VRAM for the current task. Experts are parts of the model with different abilities. The goal is to reduce GPU memory requirements while preserving answer quality and speed.
01
Expert collection
Available abilities
The model has a broad set of experts — parts with different skills. A single request usually needs only a small share of them.
02
Current task
Only a portion needed
Each request engages only the experts it actually needs.
03
Working memory
A small portion in VRAM
Only that working portion sits in GPU memory. The other experts stay out of it until they are needed.
04
Useful answer
Quality and speed
The user gets a complete answer at a useful speed — on an ordinary computer rather than a powerful server.
The idea is currently being tested step by step on small experimental models, with an open research journal.
Try it
Watch the model solve a task
Choose a fictional skill and enter digits. The model writes an answer while highlighting shows the group selected during generation.
Open the live demo ↗Current stage
Where we are
We test how an experimental solution learns new tasks and transfers results to unfamiliar examples. Complete answers and more complex sequences are assessed separately.
- 01Research setting
- 02Tool preparation
- 03Testing new tasks
- 04Independent confirmation
- 05Testing practical value
The latest recheck showed an improvement on new tasks: fully correct answers rose from 10 to 19 out of 56.
The next goal is multi-step tasks and a stable result across repeated trials.
The project is in the research stage: we test the main idea step by step.
Measured result
More fully correct answers
Recheck · 7 October 2026 · 56 individual tasks · independently generated complete answers
Current work
Research milestones
September 2026
Refined the study conditions following preliminary trials.
Early October
Checked the tool and prepared a repeat study.
7 October
Achieved an improvement on new tasks: 19 fully correct answers out of 56, up from 10.
Journal
Latest checks
Full journalThe check on new tasks is complete
Fully correct individual solutions in free generation increased from 10 to 19 out of 56. Complex sequences remain unresolved; no advantage over the simple comparison rule was established.
The next-stage criterion remains unmet
Further training did not meet the required criterion. Progression to the next stage was stopped.
Checking a partial correction
The correction helped on some tasks. New examples were prepared to assess transfer.
Open questions
What we test next
- Q1
Will the result hold in independent repetitions?
- Q2
How can we achieve a stable result on multi-step tasks?
- Q3
How much does the approach gain over simpler solutions?