Open Model Lab

Models

The July baseline records exact local Ollama identities for three evaluated models without inferring unverified size or quantization details.

Purpose

This page tracks model identity and known limitations for published evaluations. The July baseline records exact Ollama IDs because mutable tags such as latest are not sufficient artifact provenance.

Evaluation records, not full model cards. Parameter count, context length, license, and quantization are not inferred where they were not independently verified. No project-produced checkpoint has been published yet.

July evaluated model artifacts

Report name Provider Ollama model Ollama ID Quantization
lfm2.5-8b Liquid AI lfm2.5:8b 9cf756159fc2 unknown
gpt-oss OpenAI gpt-oss:latest 17052f91a42e unknown
qwen3.6 Qwen qwen3.6:latest 07d35212591f unknown

Post-training base model (planned, Stage 2)

An open-weights base/instruct pair small enough for LoRA training on the M4 Max, selected in Stage 2 and recorded with exact weight hashes, license, and precision. The Qwen3 family is the starting point: the SFT preparation already pinned a Qwen3-8B MLX artifact, and İzci uses Qwen3-4B-Base as its reference model, so choosing it would let the two projects' results share a base. Its base, SFT, DPO, and agent variants are evaluated in the same runtime and precision. The July Ollama runs remain a historical baseline.

Future model categories

  • Base models
  • SFT checkpoints
  • DPO checkpoints
  • Agent-evaluated variants
  • Multimodal models (Phase 2)
  • Project-trained models (Phase 2)

Model card fields

  • model name
  • source
  • license
  • size
  • context length
  • quantization
  • adapter/checkpoint
  • intended eval use
  • known limitations
  • date added