180B total6B activeQwen Community 1.0Vision + text262K native context
Fact. The official card describes an experimental vision-language architecture with 125B language-model parameters, 51B n-gram embeddings, 4B MTP parameters, and 6B activated parameters. It introduces sparse attention, claims native 262K context extensible to one million tokens, and had about 4.2K likes when observed.
Inference. The interesting claim is not “180B,” but whether 6B activation plus sparse attention lowers real serving cost without moving the bottleneck into memory, routing, or framework support. Qwen’s coding, office-agent, and computer-use tables justify an API trial, but several comparisons use maker-controlled harnesses, corrected tasks, in-house benchmarks, or model-based judges.
Fact. Z.ai’s official card describes a newly trained 320B-parameter multimodal model with 18B active parameters, hybrid sparse and linear attention, FP8 weights, adjustable reasoning effort, and support in Transformers, vLLM, SGLang, TokenSpeed, KTransformers, and Unsloth.
Inference. MIT licensing makes this the cleaner commercial experiment, but not a casual local model. The card’s “one-tenth the price” and near-frontier coding claims refer to the vendor’s service and evaluation setup; benchmark reproduction may require very long contexts, six-hour timeouts, maximum reasoning, and model judges. Compare end-to-end task cost rather than token price or active parameters alone.
03
A real multimodal release with a data-centre footprint
305B parametersMITExperimentalVision + text agents~18K monthly downloads
Fact. DeepSeek’s official card calls this the first experimental multimodal model in the V4 family. The repository is MIT-licensed, lists a 305B-parameter checkpoint, provides reference prompt encoding and inference code, and reports about 18K downloads and 448 likes when observed. Its vLLM recipe targets one node with four GB300 GPUs.
Inference. This is meaningful because an official, permissively licensed multimodal agent model has appeared—not because the vendor table settles its quality. The reported agent scores use DeepSeek’s own harness, maximum reasoning effort and selected sampling settings; several comparisons are close enough that deployment cost, tool reliability and visual failure modes matter more than a few benchmark points.
How this page is updated
Each review reads Hugging Face’s official trending page and the linked model cards. It records the observation date, separates repository facts from my interpretation, and avoids executing untrusted model code or downloading weights. Trending rank is volatile and is never treated as a quality score.