Now, next, planned and research

The model KOLDOS runs today.

Questions, explanations and clarifications about your code and your project.

Fixing bugs, writing features, refactoring and answering questions about a codebase.

Working with no per-request cost and no prompts sent to a model server.

Acting on the project instead of guessing, and checking results against the real files.

Not repeating the same rules, and warning before a mistake that already happened once.

Reading error screenshots, diagrams and charts during a coding task.

Current versions, documentation and facts newer than the model.

Working with logs, documents and code that is not in the project yet.

Editor extensions, scripts and bots running on the same computer.

Consolidating repeated project memories into long-term memory, and promoting lessons from past failures. Designed; waiting for real use to calibrate.

Finishing multi-step work reliably inside one turn, and resuming work a turn left pending.

A comparable history of measurements across versions, built on the existing evaluation harnesses.

Using KOLDOS from apps, servers and Discord bots without local hardware.

Tasks that need to touch live systems, with permissions granted per domain and per plan.

Large tasks divided among planned sub-executions, with testers and reviewers.

Smaller helper models for specific steps of a turn.

Linux is in scope and macOS best-effort. Windows 11 comes first.

Publishing KOLDOS with versioned bundles: system, model and evaluation suite validated together.

A structured reasoning process that plans, researches, verifies, criticizes and rethinks before answering. Planned and experimental; not available.

Search Web, Deep Web, Terminal Test and Web Tools. Planned; not available.

Possible future generation, in planning.

Fine-tuning KOLDOS 7B on annotated real-use data. On hold until there is enough of it; not training is a possible outcome.

Research into quantization, CPU/GPU offloading, KV cache, optimized kernels and weight formats so a model of about 22B parameters can run in a practical way with 8 GB of VRAM. Not a promise that it fits entirely in 8 GB.

Splitting cognition into more specialized modules. Not started.