Cheva Labs Logo
Cheva Labsweb

Lab Experiments & Research Notes

Public notebook containing hypothesis tests, memory leak evaluations, quantization degradation curves, and hardware thermal logs.

Filter Topic:

Chronological Research Record

All experiments follow a strict hypothesis-driven structure with public reproduction commands.

2026

Q3 2026 Active Test Cycle3 Studies Published
LAB-103
2026Ongoing

TamaBench — Small Models, Long Horizons

A lightweight benchmark that puts a small or local model in a persistent virtual-pet sandbox. The agent must plan across three simulated days, use structured tools, manage money and supplies, and recover when delayed consequences go wrong.

Hypothesis: A focused automation harness can extract more useful autonomy from a small model by reducing unnecessary wake-ups, tool calls, tokens, and compute without changing the model itself.
Category: AgentsCredit: Pichetpong Muangsiri
SourceView Note
LAB-102
2026Ongoing

EV Terminal — A Local AI Agent on Android

A small Android agent that loads Qwen3.5-0.8B only when a task needs model intelligence, routes deterministic work through explicit tools, and releases model resources when the task ends. It runs without an Ollama server or an always-on cloud backend.

Hypothesis: A small on-device model can still be useful when the harness owns routing, tool execution, and task lifecycle instead of asking the model to do every operation itself.
Category: AgentsCredit: Cheva1234
SourceView Note
LAB-101
2026In Review

Local LLM Inference Under Limited VRAM — MoE vs Dense vs RAM Offloading

Exploratory study of Mixture-of-Experts, dense models, and RAM offloading on an NVIDIA RTX 3060 12 GB + 16 GB DDR4. Tests whether sparse MoE can use system RAM as a capacity tier: Gemma 4 26B-A4B stays interactive (12–19 tok/s at 16K), while the dense Mistral Small 3.1 24B collapses to 3.35 tok/s and Qwen-AgentWorld 35B-A3B reaches double digits only with high latency and memory pressure.

Hypothesis: A sparse MoE with low active parameters can use system RAM as a model-capacity tier more effectively than a dense architecture under GPU/RAM hybrid inference, so total parameter count alone does not predict throughput.
Category: Models
View Note

Community Lab

Put good work on the same notebook.

Share a public project, benchmark, model, or experiment. Cheva Labs keeps the original source attached, names the contributor, and writes an independent note around the work.

Credit stays with the maker

Every entry shows the author, platform, and original link. A Lab note is a guide to the work—not a claim of ownership.

GitHubRepositories, benchmarks, and reproducible experiments.Hugging FaceModels, datasets, and evaluation artifacts.