Check the specific checkpoint
FrontiersMind’s October 29 Lumma-Fev v2 announcement includes released 4B and 27B decision models. They score predefined options rather than generate chat prose, and their repositories list Apache 2.0 licensing. The 0.6B v2 remains upcoming. [1][2][3]
For 4B and 27B, full-attention layers become bidirectional and lumma-fev 0.3 or later is required. The 4B card recommends about 10 GB of GPU memory in BF16; this is vendor guidance. [2][3]
Our October 30 source check also found actual 0.1B-v2 weights. Its card instead specifies causal attention, 2,048-token context, and SDK 0.4 or later, with CPU/small-GPU operation claimed by the publisher. Current PyPI metadata confirms 0.4 release files. Older upcoming-language in the blog is stale. [4][5]
Read the evaluation boundary
The 27B card says five of eight benchmark tasks are in-distribution because their training splits were used. The tiny card also reports weak reasoning and judgment results. Neither establishes broad performance on your unseen tasks. [3][5]
Our take
For AI geeks and experienced technicians, choose a checkpoint for a specific decision task, then evaluate held-out examples. Small files and tidy probabilities do not establish sound judgment. We inspected sources and file listings without downloading or running weights.