Separate the model decision from the workflow decision
Before comparing models, define the task: drafting a record summary, answering from approved documents or proposing a structured update. Agree how the result will be checked and which actions require approval. Both managed and self-hosted models need those application controls.
This guide compares deployment choices. It does not present a Growbiz benchmark or a claim that one model is universally cheaper or more accurate. Model versions, hosting conditions and the difficulty of the task can change the result.
What the Qwen model name tells you
Qwen’s official model card lists Qwen3.6-35B-A3B as a mixture-of-experts model with 35 billion total parameters and 3 billion activated parameters, under the Apache 2.0 licence. The activated count describes computation for a token; it does not mean that the full deployment needs memory for only a 3-billion-parameter model.
Qwen: Qwen3.6-35B-A3B model card
Estimate hardware from the selected weights, numerical precision, serving framework, context usage and expected concurrency. Include memory used by the serving process and request cache. Validate with the intended workload before committing to infrastructure.
| Consideration | Managed API | Self-hosted model |
|---|---|---|
| Infrastructure | Provider operates model-serving infrastructure; your team operates the integration. | Your team or hosting partner operates capacity, updates, availability and recovery. |
| Cost | Usage and feature charges, plus integration and review effort. | Compute, storage and operations, plus integration and review effort. |
| Data flow | Check the selected provider, endpoint, contract and application logging. | Check the full serving environment, dependencies, logs, backups and administrative access. |
| Model changes | Track provider version support and test migrations. | Choose and maintain weights, serving software and tested configurations. |
| Quality | Evaluate the exact model and settings on representative tasks. | Evaluate the exact weights, precision and serving settings on the same tasks. |
Confirm which Salesforce capability you are extending
Choosing Agentforce’s reasoning model and calling a model from a custom action are different decisions. Salesforce documents supported model options for the reasoning engine. A custom action or prompt-based feature can have its own model configuration; that does not automatically replace the engine behind the agent.
Salesforce: select an Agentforce model option
Salesforce also documents supported models and bring-your-own-model options for specific generative AI features. Check the current feature, edition, region and integration route. Do not assume that a self-hosted endpoint will work in every Agentforce or Prompt Builder configuration without validation.
Salesforce: supported large language models
Build a small, representative evaluation set
- Use approved examples that reflect normal, incomplete and contradictory records.
- Keep the available source material, output requirements and permission checks consistent.
- Have reviewers assess factual support, omissions and correction effort without relying on a model’s own confidence score.
- Test structured responses, tool failures and attempts to bypass the allowed task.
- Measure latency and throughput at the concurrency you expect to operate.
- Record the exact model version and configuration so the comparison can be repeated.
Retrieve only context the user may access. A larger context window does not remove record permissions or make outdated documents current. A smaller model may suit a bounded task, but that is a result to establish, not an assumption from its size.
Compare cost per accepted task
Include model or compute charges, failed attempts, retries, monitoring, maintenance and human review. Divide by work accepted under the agreed criteria. Cheap outputs that need extensive correction may cost more to use. Self-hosting removes a particular API bill, not the cost of running the service.
Anthropic: current model and feature pricing
Data residency and compliance depend on the complete implementation and the organisation’s requirements. Hosting weights yourself does not establish either by itself. Make the final choice with the people responsible for the workflow, security and ongoing operation, using the pilot evidence and full cost estimate.
Discuss a similar workflow
Bring a description of the process, the systems involved and one example of where work gets stuck. We can use discovery to establish the requirements, dependencies and next step before preparing a solution and estimate. Please leave private customer, applicant and patient records out of the website enquiry.