Are pretrained TTT checkpoints (Qwen3-8B / LLaMA-3.1-8B) released or planned?

Hi, congrats on the ICLR 2026 oral.

I'm trying to use this codebase to study whether the In-Place TTT fast-weight mechanism can act as a conversational-memory substrate (ingest a dialogue, drop the context, query the model — does the fast weight retain the facts?). I've read through the repo and want to confirm a couple of things before committing GPU spend.

**1. Are pretrained TTT checkpoints publicly available?**
I checked the `ByteDance-Seed` HF org and didn't see any model with `ttt` / `place` / `test` in the name. `eval_config/models.py` is a `("your_model", "/path/to/your_hf_model")` stub, and `configs/pretrain/qwen3_longct.yaml` says `model_path: /path/to/your_base_model`, which suggests users are expected to bring their own continual-pretrained checkpoint.

If a checkpoint is released somewhere, a pointer would save a lot of compute. If not, is a release planned?

**2. If we self-train at smaller scale, does the mechanism still function?**
The full recipe is `max_steps: 5000`, `global_batch_size: 64`, `max_seq_len: 65536` — many H100-days. For our purposes (a downstream evaluation, not paper-quality numbers), we're considering a much shorter run (e.g. 300–500 steps, smaller batch, seq 8192–16384) just to bring `ttt_proj` and `ttt_conv` out of init.

Is there a minimum schedule below which you'd expect the TTT layers to be effectively untrained / non-functional? Any rule of thumb on what step / loss range the fast-weight dynamics start to do something useful?

**3. Disabling TTT at inference**
For an A/B comparison we want a clean "TTT off" path. From reading `inference_model/hf_qwen3/modeling_qwen3.py` it looks like setting `config.ttt_mode = False` (or emptying `config.ttt_layers`) achieves this — is that the intended toggle, or is there a cleaner switch we missed?

Thanks!

Review eval_config/models.py, configs/pretrain/qwen3_longct.yaml, and inference_model/hf_qwen3/modeling_qwen3.py, starting with the documented model paths and inference configuration. Done means providing authoritative answers about checkpoint availability, reduced-scale training expectations, and the supported way to disable TTT at inference.

Written by the indexing model from the issue text.