Y-QB-2B-Omini β€” text development preview

This is an experimental text-only post-trained Y checkpoint, not a completed audio/visual omni model. The project name is retained for the planned full model.

Current demo selection: r21-clean. Local continuous review: external responses 13/25, memory channel 17/25, complete trajectories 0/8. Behavioral acceptance: NOT PASSED. Protocol submission success is not semantic correctness. These are repeatedly used development holdouts, not an independent blind benchmark. The model may lose constraints, misattribute actions, confuse plans and results, or fail to encode memory.

Y-authored adapter weights and this model documentation are released under MIT. The nonquantized BF16 base is Qwen/Qwen3.5-2B-Base at revision b1485b2fa6dfa1287294f269f5fb618e03d52d7c, separately distributed under its original Apache 2.0 license. Adapter tensors use FP32. No base weights are included here. NUS SMS corpus attribution is in NOTICE.md; this model license does not relicense the training corpus.

Files under checkpoints/r21-clean/ are safetensors containing the exact named trainable LoRA tensors (rank 8, alpha 16), with separate external and memory routes. They are not a standalone base model or a standard PEFT adapter directory. release.json records the target modules, immutable hashes and full local review. The matching loader and inference protocol are in the Space source.

The demo uses the existing C# Harness for scope validation, memory commits, interruption and forgetting. Current perception and short lived dialogue are bounded; deleted history is not an inference retrieval source. Tests and training run locally.

Latest local training and review status β€” updated independently of the demo selection.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for MuXue1230/Y-QB-2B-Omini

Adapter
(47)
this model