Qwen and NVIDIA Quietly Publish New Model Repos on Hugging Face, Details Sparse

Three new model repositories appeared on Hugging Face in quick succession this week: Qwen/Qwen3.8-2.4T-A95B, an FP8-quantized sibling Qwen/Qwen3.8-2.4T-A95B-FP8, and nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16. All three listings are live on the Hugging Face Hub, but as of publication, neither Qwen (Alibaba) nor NVIDIA has issued a blog post, technical report, or benchmark release to accompany them.

What the naming suggests

Hugging Face repo names for Qwen and NVIDIA model families typically encode architecture details, and these follow that pattern — though none of it is independently confirmed:

• Qwen3.8-2.4T-A95B: the "2.4T" segment implies a total parameter count near 2.4 trillion, and "A95B" implies roughly 95 billion active parameters per forward pass, consistent with a sparse mixture-of-experts (MoE) design. The FP8 variant is presumably a quantized release of the same checkpoint for lower-memory inference.
• Nemotron-3.5-Lightning-30B-A3B-BF16: the naming implies a 30-billion total-parameter MoE model with roughly 3 billion active parameters, released in BF16 precision, and branded "Lightning" — suggesting NVIDIA is targeting low-latency or low-cost inference rather than peak capability.

These are inferences drawn from established Qwen and Nemotron naming conventions, not confirmed specifications. No model card text, README, or config file details were included in the source material, so context window length, training data cutoff, license terms, and pricing remain undisclosed.

No benchmarks, no pricing, no confirmation

None of the three listings come with published benchmark scores (no MMLU, HumanEval, or MT-Bench figures were provided), and no company statement accompanies the repos. There is no confirmation of a training cutoff date, no stated context window, and no pricing information for API access — if API access is even planned. It's possible these are staged or partial uploads ahead of a formal announcement, a common practice on Hugging Face where weights sometimes go live before a coordinated release post.

Given the size implied by "2.4T" total parameters, a Qwen3.8 release at that scale would represent one of the largest openly-listed weight sets to date if confirmed — but until Alibaba's Qwen team publishes documentation, that figure should be treated as speculative, derived only from the repo name.

What this means

This is a roundup of repository listings, not a confirmed product launch. Multiple outlets picked up the same bare Hugging Face URLs because model repos appearing without accompanying documentation often signal an imminent announcement — labs frequently push weights ahead of blog posts to let early testers begin downloading. Until Qwen and NVIDIA publish technical reports, benchmark tables, and licensing terms, treat parameter counts, active-parameter figures, and capability claims implied by the names as unverified. Readers building on these models should wait for official model cards before making architecture or deployment assumptions, and should not assume API pricing or availability until either company confirms it directly.