Holo4 is not one interchangeable commercial deployment option. H Company has released a computer-use model family whose 35B-A3B weights carry Apache 2.0, while the 27B weights carry CC BY-NC 4.0. For an enterprise considering local automation, the first decision is therefore which artifact and usage rights fit the workload—not which benchmark score is higher. Hosted API access is a separate contractual question; a published token price does not settle the rights for your intended use.
What changed with the Holo4 release
In its announcement dated 28 September 2026, H Company introduced Holo4 27B dense and Holo4 35B-A3B mixture-of-experts models, alongside Holotron4 Nano. The company says both Holo4 sizes are available through its Models API and offers downloadable weights, including quantized formats. Optimized DSpark drafter checkpoints are described as coming in the following days, not as an already verified deployment dependency.
The models are designed to combine screenshots, code execution, MCP and API tools across desktop, web and mobile tasks. That is relevant where a business process crosses a legacy interface and a newer service API. It does not establish that a model can safely use your ERP session, or that every requested interface is available in your chosen runtime. The model card describes a harness that supplies screenshots and tool results, executes requested actions, and returns their outcomes. The harness and its permissions remain part of the system you must evaluate.
This article examines the release documentation, not a hands-on benchmark. The distinct buyer question is whether the particular Holo4 variant you can license and operate also passes your workflow acceptance tests.
The weights have different licenses
The Holo4-27B model card identifies its weights as CC BY-NC 4.0. The Creative Commons deed explicitly includes a NonCommercial restriction and warns that the deed is only a summary of the legal terms. H Company's launch page labels this variant “Research only.” Do not treat a downloadable checkpoint as blanket permission for commercial self-hosting. Have counsel assess the proposed use and obtain additional permission where required; this is procurement guidance, not legal advice.
The Holo4-35B-A3B model card instead identifies Apache License 2.0. This is a materially different starting point for commercial evaluation, but not a certification of the complete application. Keep the checkpoint revision, associated license and notices, serving dependencies, modifications and intended use together. A quantized derivative is a separate artifact to inspect, not a reason to assume that obligations disappeared.
Parameter count is also a poor shortcut here: 27B is dense; 35B-A3B is a mixture-of-experts variant with fewer active parameters than its total count. Do not infer memory fit, throughput or equivalent task quality from the model names. Published configuration limits are not measured capacity under your screenshot volume and concurrent sessions. The broader model procurement control explains why workload, hosting and change policy must be assessed together.
Hosted availability and prices: what is actually documented
The launch page lists prices in US dollars per million tokens. For Holo4-35B-A3B it gives $0.30 input, $0.03 cached input and $2.00 output. For Holo4-27B it gives $0.40 input, $0.04 cached input and $3.00 output. These are the vendor's published API rates at retrieval, not a quote for a complete automated business transaction.
There is a documentation boundary to preserve. The Models API introduction retrieved for this article still recommends Holo3.1 and describes that generation's availability. It states a default zero-retention policy for prompts and responses, but this article does not turn that generic page into a verified Holo4-specific regional or contractual guarantee. Confirm the exact callable model ID, account eligibility, service terms, retention scope and processing location with the supplier before sending enterprise data. The announcement does not establish EU-only processing.
For the 27B variant in particular, ask explicitly how the hosted offer and its “Research only” label apply to the proposed commercial workload. Do not infer that the noncommercial weights license necessarily defines every hosted-service permission, or that paying for tokens automatically grants those permissions. Record the supplier's answer and the reviewed contract before approval.
Why the stronger headline score cannot choose your deployment
H Company reports an OSWorld 2.0 score of 61.7% for Holo4 27B and 30.9% for 35B-A3B, with respective token-based costs of $1.22 and $0.61 per task. These are vendor-reported benchmark measurements, not independent comparisons or production success guarantees. The announcement distinguishes partial score from success rate and reports a single run for this benchmark. Releases, harnesses, effort settings and task subsets differ across the external baselines.
The AutomationBench comparison has another important qualification: H says 480 of its 600 public evaluation tasks fall in the split from which it collected training data; it separately reports the 120 held-out tasks and says official private-set evaluation is still pending. That disclosure is useful, but it rules out treating the public headline as an uncontaminated estimate of performance on unseen company workflows.
For a pilot, keep a held-out task set outside prompt and harness tuning. Verify final application state, not the agent's completion message. Record failed attempts, intervention time, wall-clock duration, input and output tokens, and the expense of review and recovery. A lower cost per attempted task can still produce a more expensive accepted outcome. The existing production agent evaluation guide provides the measurement foundation; Holo4 adds a variant-specific licensing filter before that comparison even starts.
A five-checkpoint decision checklist
1. Identify the exact artifact and permitted use
Capture repository, revision, weight format, license file and intended business process. For 27B, stop commercial self-hosting approval until the noncommercial restriction has been reviewed and suitable rights established. For 35B-A3B, retain the Apache notices and review the complete distribution. Evidence required: a versioned artifact record and an accountable legal or procurement decision, not a family-level “open weights” label.
2. Separate hosted approval from local approval
For hosted inference, confirm model ID, access, rates, contract and data path. For local inference, test the actual hardware, concurrency and runtime you will support. A local model can still send business data through external tools. Evidence required: a data-flow inventory covering screenshots, prompts, tool responses, logs and every destination—not merely the location of the GPU.
3. Evaluate the model and harness as one release
Pin the serving configuration, prompt, tool schemas and harness revision. Exercise the same held-out tasks through each eligible candidate. Include German interface labels, modal dialogs, permission failures and tasks that require switching between GUI and API access. Evidence required: reproducible results with task-level outcomes and operator interventions. These are proposed acceptance tests, not capabilities independently demonstrated here.
4. Prevent an interface switch from widening authority
Use narrowly scoped test accounts, disposable desktop or browser environments, restricted network access and bounded execution. A denied API action must not become permitted because the agent can click the corresponding button. Keep irreversible writes behind approval enforced outside model text. The browser-agent control boundary supplies the baseline; combining interfaces makes cross-channel consistency the test to add.
5. Approve a bounded pilot, not the model family
Name the owner, allowed tasks, budget, stop conditions and fallback. Preserve the evaluation record and rights decision for that exact variant. Re-evaluate when the checkpoint, license, harness or service behavior changes. If the Apache-licensed variant fails your workload, do not silently substitute the higher-scoring noncommercial checkpoint. Reopen the technical and procurement decision together.
The next useful step
Holo4 makes a concrete distinction visible: downloadable weights, benchmark capability and approved commercial operation are three different facts. A defensible pilot connects them with artifact-level evidence rather than a general claim that the family is open.
If you need to turn that distinction into a deployable system, my AI automation implementation services can help define the evaluation harness, tool permissions and operating acceptance criteria. Start with one reversible workflow and an explicitly eligible model; expand only after the measured result and the reviewed deployment terms support it.
