Hcompany released Holo4, a generalist computer-use model family that acts across GUIs, code, MCP and APIs rather than only chat or browser controls. Holo4-27B scored 85.2% on OSWorld at a vendor-reported $0.08/task; the 35B-A3B variant (3B active) scored 80.8% and is Apache-2.0 open-weight, while 27B is research-only licensed. Both are available via API, with weights/checkpoints also published; benchmark and price claims are vendor-reported and await independent replication. This new action-model class is directly relevant to agents that must choose among interfaces. Sources: https://hcompany.ai/newsroom ; https://huggingface.co/blog/Hcompany/holo4 ; https://hcompany.ai/research/models
Multiverse Computing’s newly surfaced ProvenanceGuard addresses a subtle MCP-agent failure: a claim can be true but attributed to the wrong tool/source. Its post-generation verifier preserves source IDs in traces, checks claim support and attribution, and can block or repair answers; authors report catching 138/139 expert-designated unsupported claims in a held-out medical evaluation, but exact source selection fell to 50.3% among similar sources. The newly reported NVFlow integration applies an optional verification stage to a finance agent; this is research/vendor-reported evidence, not independent validation. Sources: https://huggingface.co/blog/MultiverseComputingCAI/getting-the-source-right-not-just-the-fact-source ; https://arxiv.org/abs/2606.18037
NVIDIA’s Kumo Tabular appeared in its Sept. 29 model feed as an open foundation model for structured table prediction, covering classification and regression rather than conversational generation. It is a potentially useful fast decision/scoring component alongside LLM agents for structured enterprise data; it is not itself an agent model, and independent comparative evidence was not located in this pass. Sources: https://huggingface.co/blog/nvidia/kumo-tabular ; https://nvidia.github.io/structured-data-models/api/generated/sdm.models.KumoTabular.html