Start with a hosted API model pinned to a version, and move to open weights when a contract, a regulator or your monthly bill gives you a concrete reason. Open models now trail closed ones only slightly on some benchmarks, so the choice turns on data control, volume and who runs the servers.
Start with an API, and move to open weights only when you have a concrete reason. For most small and mid-sized businesses in India, a hosted model from Anthropic, OpenAI or Google, pinned to a version, is the fastest and cheapest way to prove an AI workflow. Open-weight models become the right call when a client contract, a regulator or your own risk rules require data to stay on infrastructure you control, or when volume makes per-token pricing your biggest AI cost.
The debate generates more heat than the decision deserves. It is a build-versus-buy question wearing a philosophy costume. The quality gap that once settled the argument has narrowed sharply, so the choice now turns on data control, cost at volume and who keeps the servers running.
- gap between open-weight and closed models on some benchmarks, down from 8% in one year
- 1.7%
- gap between open-weight and closed models on some benchmarks, down from 8% in one year
- of memory is enough to run OpenAI's open-weight gpt-oss-20b
- 16GB
- of memory is enough to run OpenAI's open-weight gpt-oss-20b
- a single data-centre GPU of this size fits the larger gpt-oss-120b
- 80GB
- a single data-centre GPU of this size fits the larger gpt-oss-120b
- Anthropic's price multiplier for US-only processing, the only alternative to global routing
- 1.1x
- Anthropic's price multiplier for US-only processing, the only alternative to global routing
01
What open weights means, and what it does not
An open-weight model is one whose trained weights you can download and run on your own hardware or on a cloud server you choose. An API model stays on the provider's servers: you send text and pay per token. Open weights is not always open source, and the licence matters more than the label.
OpenAI's gpt-oss models, released in August 2025, and Alibaba's Qwen3, which covers more than 100 languages and dialects, use the permissive Apache 2.0 licence. Meta's Llama 4 uses its own community licence: companies with more than 700 million monthly active users must ask Meta for a separate licence, and anyone distributing it must display Built with Llama.
Language support is a practical check for Kerala. Meta lists 12 supported languages for Llama 4, including Hindi but not Malayalam. Sarvam-M, a 24 billion parameter open model built on Mistral Small and released under Apache 2.0, was post-trained on Indian languages alongside English, and its makers report a 20% average gain on Indian-language benchmarks over the base model. Its model card lists English and ten Indian languages, Malayalam among them, so it is worth a test on your own customers' messages.
02
Capability: the gap is small, but not zero
Stanford's 2025 AI Index found that open-weight models cut their deficit to closed models on some benchmarks from 8% to 1.7% in a single year. For routine business tasks, such as sorting enquiries, pulling fields out of invoices or drafting standard replies, a good open model is often enough, though only a test on your own examples will tell you.
The frontier keeps moving, though, and APIs get there first. Since the start of February, Anthropic has released Claude Opus 4.6 and Sonnet 4.6, Google has previewed Gemini 3.1 Pro and OpenAI has launched GPT-5.4 with a 1 million token context window. With an API, a newer model is a change of model name away, though you should still test it before you switch. With open weights, every upgrade is your job.
03
Cost: paying per token versus paying for servers
API prices are quoted per million tokens, in US dollars. Anthropic charges $3 per million input tokens and $15 per million output tokens for Claude Sonnet 4.6, and $1 and $5 for the smaller Haiku 4.5, with half price for jobs sent through its Batch API.
A hypothetical example: a logistics firm in Kochi summarises 3,000 customer emails a month, each using 2,000 tokens in and 300 tokens out. That is 6 million input and 0.9 million output tokens, about $10.50 a month on Haiku 4.5 or $31.50 on Sonnet 4.6. A server costs the same whether it handles 30 requests or 30,000, so at small-business volumes the API is usually far cheaper.
Even so, self-hosting has become far more accessible. OpenAI says gpt-oss-20b runs within 16GB of memory and gpt-oss-120b fits on a single 80GB GPU such as an NVIDIA H100, and the AI Index found AI hardware costs falling about 30% a year. The model is free. The server, the power and the person who runs it are not.
| Question | Hosted API model | Open-weight model |
|---|---|---|
| Where does data go? | The provider's servers, in the regions it offers | Wherever you run it, including a server in India |
| What do you pay for? | Tokens used, billed in US dollars | Hardware or GPU rental, power and staff time |
| Who upgrades it? | The provider, on its schedule | You, when you choose |
| Effort to start | An account and an API key | A server, monitoring and security patching |
| Main risk | Changes to price, policy or model availability | Licence terms and your own upkeep |
04
Data location and the Indian rules
For most businesses, Indian law does not force AI processing to stay in the country. Under the Digital Personal Data Protection Act, 2023, personal data can be transferred abroad except to countries the central government restricts by notification. Sector rules can be stricter: the Reserve Bank of India requires payment system operators to store payment system data only in India. Client contracts can add conditions of their own.
Provider options are narrower than many owners assume. When Anthropic introduced data residency controls for its own API on 5 February 2026, processing on its newest models could be pinned to the US, at 1.1 times the normal price, or left global; there was no Indian option. Cloud platforms such as Amazon Bedrock and Google Cloud set their own regional endpoints, so check those separately. So if a contract says data must stay in India, a hosted API may not satisfy it, while an open-weight model on a server in India can.
05
A rule that survives contact
Choose open weights when one of three things is true: a regulator, client contract or your own risk posture forbids data leaving your control; your volume makes per-token pricing the dominant cost; or you need the model to work offline, or never to change without your say. Otherwise start with an API, version-pinned, behind a thin layer in your code that lets you switch later.
For a typical small business the sequence is the same: API first, prove the workflow, then re-evaluate hosting when the invoice or the compliance question actually arrives. Most early wins come from connecting the tools you already pay for, not from where the model runs.
06
What to do next
- List the data each AI workflow will touch, and note any contract or regulator that limits where it can go.
- Prototype on a hosted API with a pinned model version and only the personal data the task needs.
- Route every model call through one function in your code, so moving to another provider or to open weights is a small change.
- Track monthly token spend, and revisit self-hosting when the bill or a compliance requirement justifies a server and the person to run it.
- If you test open models, check the licence and the supported-language list against your customers, including Malayalam.
- Keep a human approval step for anything a customer or regulator will see.
Sources
Facts current as of 12 March 2026. Prices are list prices in US dollars, and licences and data residency options change, so check each provider's current terms.