Skip to content

PROJECT_TRUTH.md

Latest commit

 

History

History
65 lines (50 loc) · 3.21 KB

PROJECT_TRUTH.md

File metadata and controls

65 lines (50 loc) · 3.21 KB
Raw

Project Truth — What Kimari Is and Isn't

This document exists to prevent confusion, inflated expectations, and unfair criticism.

What Kimari Is

  • A local AI framework and CLI for running language models on consumer NVIDIA GPUs.
  • A GGUF/llama.cpp workflow designed for older GPUs (GTX 1060, GTX 1080).
  • A local OpenAI-compatible endpoint helper that manages llama-server lifecycle.
  • A Gateway Dashboard for monitoring and managing the local AI environment.
  • An integration tool for Open WebUI, Continue.dev, OpenClaw, and Hermes.
  • Open-source (MIT license). Python + TypeScript/Next.js stack.
  • Local-first. No cloud dependency. No telemetry. No subscriptions.

What Kimari Is NOT

  • Not a new inference engine. Kimari uses llama.cpp under the hood. It organizes, diagnoses, configures, profiles, documents and integrates llama.cpp for hardware-limited users.
  • Not a public Kimari-4B model. No Kimari-4B weights, adapters or GGUF files are publicly available. The gate is BLOCKED.
  • Not a production server. This is alpha software. Useful today, but not production-ready.
  • Not a benchmark leaderboard. No public benchmark claims are made. Performance numbers shown in documentation are validated test results (TinyLlama), not Kimari-4B benchmarks.
  • Not a replacement for Ollama, LM Studio or text-generation-webui. Kimari is a complementary tool focused on older GPU optimization and honest alpha development.

Current State (v0.1.87-alpha)

Area Status
Framework / CLI ✅ Usable alpha
Local GGUF runtime ✅ Working (with llama-server)
OpenAI-compatible endpoint ✅ Working
GTX 1060 validation ✅ Validated with TinyLlama test model
Gateway Dashboard ✅ Local preview (127.0.0.1:3105)
One-command install ✅ install.sh / install.ps1
Private adapter experiments 🔒 Private (SFT v2 on SmolLM3-3B)
Private manual review 🔒 Completed (safety_fix_required)
Public Kimari-4B weights ❌ Not released
Public GGUF Kimari model ❌ Not released
Public benchmark claims ❌ None
Release gate 🔒 BLOCKED

Label Legend

Throughout Kimari documentation and the GitHub Pages site, you will see data labeled as:

Label Meaning
VALIDATED Measured result from real hardware, reproducible
ESTIMATED Reasoned projection based on similar hardware/models
SIMULATED UI Visual demo without real data backend
PLANNED Feature exists only in roadmap
NOT RELEASED Exists privately but not available publicly

When in doubt about a claim, assume it is NOT VALIDATED unless explicitly marked VALIDATED.

Safety Commitments

  • No telemetry. No phone home. No analytics on user machines.
  • Default bind is 127.0.0.1. Non-local bind requires explicit flag.
  • No public Kimari-4B weights until safety regressions are fixed.
  • No benchmark claims without reproducible validation.
  • No GGUF export until manual review passes without safety issues.
  • Model registry hashes will be pinned before models reach recommended status.

This document is a living commitment to honesty. If any claim becomes inaccurate, this file will be updated before the next public commit.