Limitations and attribution
Research release with very limited evaluation. Identity claims describe the customization and do not change the original model authorship. See evaluation.json for actual generated answers. Do not infer benchmark quality from successful name responses.
Training sources include prefixes of SmolLM-Corpus (ODC-BY), permissively filtered CodeParrot, Smol-SmolTalk (Apache-2.0), and self-oss-instruct-sc2-exec-filter-50k (ODC-BY). See provenance files where supplied. Original dataset notices and licenses apply to the source data; the training datasets are not redistributed here. No affiliation with Qwen or dataset authors is implied.
Full merged Safetensors
All original parameters and experts are present. Only the targeted linear weights receive the learned LoRA deltas. The base model is not retrained from scratch. Load the repository directly with its original Transformers model class; the adapter loading example above is an alternative using subfolder="adapter".
Load the full merged weights directly
These weights already include the learned adapter. Do not apply the adapter again when loading this full release. Transformers 5.3.0 or a compatible newer version is required for the Qwen architectures.
import torchfrom transformers import Qwen3_5MoeForConditionalGeneration, AutoTokenizerrepo = "Convergent-ch/Alpina-Large-2.0-Mega-122B"model = Qwen3_5MoeForConditionalGeneration.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")tokenizer = AutoTokenizer.from_pretrained(repo)messages = [{"role":"system","content":"You are a helpful assistant. Answer in English."},{"role":"user","content":"Who are you?"}]inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, enable_thinking=False, return_tensors="pt", return_dict=True).to(model.device)output = model.generate(**inputs, max_new_tokens=96, do_sample=False)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
The full tensor payload is 250.2 GB (233.0 GiB), before runtime memory. Stored parameter count: 125,086,497,008. Counts include all tensors in the original base checkpoint, including auxiliary components where present. The published base name is its upstream nominal size, which can differ from the complete stored-tensor count.