Keural's Qwen experiment lands as a 78.8GB GGUF

Keural-Lab-1031 gives local-model experimenters a new llama.cpp download, with a refreshingly explicit scorecard rather than a claim to beat everything else. The community developer mkd-minkeun describes the checkpoint as an experiment based on Qwen3.8-Flash-Next. Its original weights were uploaded on November 2 at 12:45:44 UTC, and the GGUF artifact followed at 12:47 UTC.

What the download includes

The GGUF repository contains Keural-Lab-1031-Q4_K_M.gguf, listed at 78,832,809,056 bytes, or about 78.8 GB, plus a separate BF16 vision projector of about 0.9 GB. The English card labels the projector optional and supplies a llama-server command using --mmproj, a 32,768-token context setting and --jinja.

This is a community distribution for people who want to try the checkpoint through llama.cpp, alongside the original BF16 release’s vLLM example. Download size is not a tested RAM or accelerator-memory requirement. The reviewed card does not specify a minimum llama.cpp revision, a hardware configuration or measured generation speed, and AITrending has not run either artifact.

An experiment, with its own comparison table

The original checkpoint card reports evaluations using lm-evaluation-harness with vLLM, full test sets and thinking enabled. Its comparison model, Qwen3.6-35B-A3B, scores higher on every listed task: MMLU, MMLU-Pro, GPQA Diamond, GSM8K and IFEval. For example, the author reports GPQA Diamond scores of 0.678 for Keural and 0.803 for that comparison model.

Those are the developer’s reported measurements for the original checkpoint. They do not measure the new Q4_K_M file’s quality or establish that quantization preserves the same scores. The card does not explain the training recipe in detail.

There is also a sizing discrepancy worth resolving before planning a deployment: the card estimates roughly 65 billion total parameters, while Hugging Face’s original checkpoint metadata inventories roughly 117 billion tensor parameters. AITrending has not reconciled those counts; the file sizes above are directly listed artifacts, not an inferred model-size claim.

The supplied Qwen Community License 1.0 also matters for reuse. It requires a separate Qwen license for commercial use by businesses conducting model-as-a-service or AI work-assistant operations, with an exception for qualifying internal use. This is a downloadable experiment with specific terms, rather than an unrestricted commercial license or an official Qwen release.