Current large local model
Local multimodal
Qwen3.8 27B Vision
The larger local brain for work that needs broader reasoning and vision, sized for the unified-memory NVIDIA fleet we operate today.
Small models that learn our work.
Karti Small is a growing family of local models for agents, precise tool calling, vision, and useful work even when the internet is unavailable. Every checkpoint comes from the same practical loop: collect, verify, train, evaluate, improve.
Published models
Text first. Vision next.
The 3B program teaches the tool loop. The 4B program adds sight without giving up structured action. Both are documented in public as the work happens, including the failures that change the next run.
Active SFT program
Text + tools
Karti-Small-RSI-3B
A compact local agent model trained for exact tool choice, schema-valid arguments, policy boundaries, and finishing work through a delivery tool.
SmolLM3-3B base · model card and reproducible recipe public
v0 public · SFT underway
Vision + tools
Karti-Small-VL-4B
A small local vision-language model that can inspect an image or screen and make a structured tool call about what it sees.
Qwen3.5-4B base · runnable v0 weights and recipe public
The model system
A working fleet around them.
We route work by capability and privacy instead of asking one model to do everything. Karti Small owns the focused local lanes; larger local and cloud models add breadth and frontier coding capacity.
Tracking
Evaluation queue
DeepSeek V4 Flash Vision
A fast multimodal candidate we are tracking alongside other flash-class releases, with local deployment gated by measured fit on a single DGX Spark.
Subscribed
Cloud frontier
Codex + Claude Code
Frontier coding capacity for difficult research and implementation work. These complement the local fleet; they do not replace the owned model path.
Orchestration
Prime Agent + RLM
Prime Agent is the operating harness. The recursive language-model loop decomposes work, selects the right local or cloud capability, records evidence, and turns reviewed outcomes into the next improvement cycle.
Recursive self-improvement, made concrete.
RSI here is an engineering discipline, not a magic switch. Every cycle turns reviewed experience into a candidate, measures it against behavior we care about, and carries only useful evidence forward.
01
Collect
Useful, reviewed traces from the work we actually want the model to do.
02
Verify
Rights, privacy, structure, and behavior checks before a row can teach the model.
03
Train
A bounded monthly checkpoint from an immutable, reproducible starting point.
04
Evaluate
Held-out tool and behavior contracts decide what improved.
05
Improve
The evidence shapes the next data mixture and the next checkpoint.
Open foundations. Focused adaptation.
SmolLM3-3B supplies the text-and-tools foundation; Qwen3.5-4B adds a native vision and tool-call path. Our work is the recurring adaptation loop around them: better data, exact tool contracts, train/serve parity, careful evaluation, and deliberate promotion from one checkpoint to the next.
Case study
From an RSI hackathon win to our first trained model.
PodMan proved the loop in a weekend. Karti-Small-RSI-3B turned the same discipline—observe, verify, improve—into a repeatable model program.