H Company released two Holo4 vision-language agents that use one harness across desktop, web and Android interfaces, code execution, MCP and conventional APIs. The dense 27B and 35B-A3B MoE models are available through an API, ungated weights, a public harness and published evaluation trajectories. H Company reports strong OSWorld results, but OSWorld 2.0 was run once and most public AutomationBench tasks came from the split used to collect training data. The 27B license is non-commercial CC BY-NC 4.0, while the 35B-A3B model uses Apache 2.0.
Availability: H Models API; ungated BF16/FP8/NVFP4/GGUF weights; public hai-agents harness; public evaluation trajectories
For computer-use teams, this is an unusually inspectable release because the weights, harness and step-by-step trajectories ship together. Trial the Apache-2.0 35B-A3B model first for commercial evaluation, even though its vendor-reported long-workflow score is weaker. Treat the 27B model as research-only without a separate commercial license, and rerun OSWorld 2.0 plus the 120 held-out AutomationBench tasks before accepting cost or quality claims.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Holo4 is available through an API, two weight releases, a public harness and evaluation trajectories; the two model licenses are CC BY-NC 4.0 and Apache 2.0.
The vendor says OSWorld 2.0 was run once, 480 of 600 public AutomationBench tasks came from the training-data collection split, and no official private-set result is available.
| Models | Holo4-27B dense based on Qwen3.8-27B; Holo4-35B-A3B MoE based on Qwen3.6-35B-A3B | ||||||
|---|---|---|---|---|---|---|---|
| Context Window | 262144 | ||||||
| Interfaces | desktop, web, Android, code execution, MCP and conventional APIs | ||||||
| Training | about 10,000 synthetic/collected tasks; supervised learning plus asynchronous online reinforcement learning | ||||||
| Availability | H Models API; ungated BF16/FP8/NVFP4/GGUF weights; public hai-agents harness; public evaluation trajectories | ||||||
| Licenses |
| ||||||
| Vendor Benchmarks |
| ||||||
| Benchmark Limits | OSWorld 2.0 and ALE-CLI are single runs; 480/600 public AutomationBench tasks are from the training-data collection split; no official private-set result; cross-model table mixes harnesses and effort levels |