Two Amharic Models, One Name, Zero Signal

What happened

Two Amharic text-generation models landed on Hugging Face two days apart in late August 2026, both naming "Qwen3.5" in their titles. One is a from-scratch 50M-parameter architecture comparison; the other is a LoRA adapter whose own model card reports a nan validation loss.

Context

The Qwen3.5 architecture family, cited in the benchmark repo to the Alibaba Qwen Team (2025/2026), gives a reference that smaller builders can re-implement at lower parameter counts. What stands out here is the pattern rather than any single model: two unconnected Amharic artefacts, published 48 hours apart, both at zero downloads and zero likes, neither backed by a published evaluation or community validation.

How it works

The 50M benchmark contains four PyTorch architectures trained from scratch on the Amharic Wikipedia corpus, all sharing an RL-trained subword tokenizer with 3,919 tokens. The Qwen3.5 Transformer variant (50.79M parameters) uses a 3:1 Gated DeltaNet/Gated Attention ratio, Zero-Centered RMSNorm, Partial RoPE at coefficient 0.25, and a Multi-Token Prediction head. HRM-Text (49.02M) uses dual-timescale H2L3 recurrence. The Mamba SSM variant (49.47M) uses a selective state-space with dt_rank=32. The Hybrid variant (52.04M) interleaves Mamba and Qwen3.5 attention at a 2:1 ratio. Weights ship as .pt checkpoint files. The model card references analysis plots but no scores appear inline.

The 4B model is a LoRA adapter (PEFT 0.20.0) on rodrigomt/Qwen3.5-4B-Uncensored-Aggressive, fine-tuned for one epoch over 15 steps at learning rate 2e-05 with AdamW-8bit. Training loss: 0.9318. Validation loss: nan. Dataset: "unknown."

Our read

The naming collision is the story. "Qwen3.5" in the 50M benchmark is an architecture family being re-implemented at research scale. "Qwen3.5-4B" in the LoRA repo is a specific checkpoint redistributed under the tag "Uncensored-Aggressive" and then fine-tuned for 15 steps. They share a name and a language. That is the connection.

The 50M benchmark is the more defensible artefact. Comparing recurrence, selective state-space, and standard attention at a matched 50M budget on a single low-resource corpus is a real question. But the card links to plots, not numbers, and the repo has zero downloads. It is a working draft, not a published result.

The 4B LoRA is harder to look at. A nan validation loss is not a hyperparameter quirk; it means the validation pass did not produce a number. The training dataset is the single word "unknown." The base model's name encodes a safety posture that no studio can audit before client delivery. And no licence appears anywhere in the card.

Two models, one language, zero community signal. Amharic open-weight NLP is still a scatter of isolated experiments.

What this changes

Nothing in the ComfyUI pipeline, the VAE stack, or any video-diffusion workflow. Both models are text-generation only.

If Amharic subtitle drafting or prompt writing enters the studio's workload, the 50M checkpoints are the only set that clears a licence check (Apache-2.0) and runs on a CPU from sub-100 MB files. They are also unvalidated research artefacts with no reported quality scores. The 4B LoRA is a legal and quality liability as published: no licence, nan validation, unknown data, an unauditable base model.

On Monday, nothing changes. Neither model enters a production path. If a quick Amharic text test is needed, the 50M Transformer checkpoint is the least-bad starting point.

License

The 50M architecture benchmark is Apache-2.0: commercial use is permitted, attribution is required. The 4B Amharic SFT LoRA has no licence stated in its model card or index. Until a licence is confirmed on both the adapter and the base model, treat it as unusable for commercial deliverables.

Key takeaways

• Two Amharic text-generation models appeared on Hugging Face on August 21 and 23, 2026, both referencing "Qwen3.5" but representing unrelated artefacts with zero downloads.
• The 50M benchmark compares four architectures at matched parameter count on Amharic Wikipedia; it is Apache-2.0, runs on CPU, but reports no inline scores.
• The 4B LoRA fine-tune carries a nan validation loss, an "unknown" dataset, no stated licence, and a base model whose name encodes an unauditable safety posture.
• Neither model touches the studio's video pipeline; both are text-generation only.
• Amharic open-weight NLP remains a collection of isolated, unvalidated experiments rather than a usable ecosystem.

Sources

1. amanuelbyte/amharic-50m-architecture-benchmark — text-generation on Hugging Face — tier 3
2. abdukuzi45/qwen3.5-4b-amharic-sft — text-generation on Hugging Face — tier 3