Qwen and LiquidAI Quietly Push New Model Weights to Hugging Face: Qwen3.8-27B, Qwen3.8-27B-FP8, and LFM2.5-VL-3B

Three New Model Repositories Surface on Hugging Face

Three separate model repositories appeared on Hugging Face in short succession: Qwen/Qwen3.8-27B, a full-precision variant; Qwen/Qwen3.8-27B-FP8, an FP8-quantized version of the same model; and LiquidAI/LFM2.5-VL-3B, a vision-language model from LiquidAI. All three listings went live without accompanying blog posts, technical reports, or benchmark disclosures from either company at the time of writing.

What's confirmed

The repository names indicate a 27-billion-parameter model from Qwen (Alibaba's model family) released in two forms — standard precision and FP8 quantization — plus a 3-billion-parameter multimodal model from LiquidAI supporting vision and language inputs, based on the "VL" designation in its name. Beyond these naming conventions, no verified details are available: Hugging Face model cards for all three repositories had not been populated with parameter breakdowns, training data cutoff dates, context window sizes, or benchmark scores as of publication.

What remains unverified

No pricing has been disclosed for any of the three models — typical for open-weight releases distributed directly through Hugging Face rather than an API. Context window length, MMLU or other standard benchmark results, and training cutoff dates are not yet published for Qwen3.8-27B, its FP8 counterpart, or LFM2.5-VL-3B. Qwen has not issued a companion announcement on its usual channels (GitHub, ModelScope, or its blog) describing architecture changes from prior Qwen releases. LiquidAI has similarly not published documentation clarifying how LFM2.5-VL-3B differs from its predecessor LFM2-VL line.

The FP8 variant of Qwen3.8-27B is presumably intended for lower-memory inference deployments, following a pattern common across the industry where labs ship both full-precision and quantized checkpoints simultaneously to serve different hardware constraints. Without an official technical report, however, the specific quantization methodology and any accuracy tradeoffs versus the full-precision Qwen3.8-27B checkpoint cannot be confirmed.

What this means

This is a case where model weights reached Hugging Face ahead of any public documentation — a pattern increasingly common among labs that treat repository uploads as a soft launch, with formal announcements following days or weeks later. Engineers evaluating these models should treat any claimed capabilities as unverified until Qwen and LiquidAI publish accompanying technical reports or benchmark suites. Until then, the safest path for builders is to run independent evaluations against existing baselines (Qwen3, LFM2-VL) rather than assume parity or improvement based on version numbering alone. Given the absence of confirmed specs, this roundup will be updated as official documentation becomes available.