The Hub is quiet on the blog side today. The activity is small and specific. A Gemma-4 mirror adds a confidence probe for disaster use. A community training diary shows how far a 139M model can go on a consumer GPU. A blog post questions how benchmark numbers turn into marketing.
On the Hub
ginigen-ai Edge-4B-TELL
ginigen-ai published Edge-4B-TELL today. It mirrors Google's Gemma-4-E4B QAT GGUF.2 The pack adds a 10 KB TELL probe for the HeliGO Android app. The base weights date from May. The release includes a Q4_0 text GGUF and a vision mmproj.1
The card lists Apache-2.0 and links to Gemma 4 terms. The probe is card-claimed. On 665 Korean disaster-procedure questions, self-reported confidence AUROC is 0.441, surface features reach 0.736, and TELL reaches 0.759 plus or minus 0.006. The gain over surface is about 0.023. Answer accuracy on the set is 27.5 percent.
The model is for offline disaster response on phones. The probe adds no generated tokens. It is meant to flag low-confidence answers for fallback.
BananaMind 2 Pro training report
BananaMind 2 Pro has been on the Hub since August 14. A community blog posted September 18 details the training run. The model has 139 million parameters. It was trained on about 100 billion tokens on an RTX 5070 Ti.3
The license is bananamind-community-license-1.0. Commercial use is capped at 200,000 MAU or $500,000 monthly revenue. The author reports 96.08 percent retention versus SmolLM2-135M on a nine-task suite.4
The author also snapshots SLM Arena. BananaMind-2-Pro is listed first with an Elo of 1163.8 on 51 battles. The sample is small.
When Benchmark Numbers Become Marketing
DedeProGames published a community audit on September 17. The post examines DavidAU and Nightmedia model cards. It focuses on a seven-task suite used as a general intelligence claim.5
The author notes ARC-C scores presented as an OpenAI or Claude zone of intelligence. Long-context labels are flagged as evaluated with a 512-token perplexity script on a training split. Tokens per second from that script is eval throughput, not generation.
The piece audits marketing language. It is not a model release.
