TL;DR: A US software developer, the closest BLS match, earns a median $135,980 a year ($11,330 a month) before benefits (May 2025); AI engineers in India working for foreign clients earn $2,690 to $10,440+ a month on our rate card. Through Second Talent, a vetted LLM developer in India starts at $25 USD an hour (160 hours for $3,999 a month), shortlisted in 24 hours.
Bengaluru's Sarvam AI published Sarvam 30B and 105B under Apache 2.0 on 6 March 2026, trained in India with a tokenizer that covers all 22 scheduled Indian languages. A developer who has served them knows what a Hindi prompt costs.
Key takeaways
- India keeps UTC+5:30 all year, so a 3:00 pm to midnight shift there overlaps four and a half hours with a 9-to-5 ET day in winter and five and a half once US daylight time starts on 14 March 2027.
- An Indian copyright assignment that names no territory is presumed to cover India only, and one that names no period lasts five years.
- India's IT Rules, amended in February 2026, require platforms to label realistic AI-generated audio, images and video.
- Notice periods for AI engineers in India run 30 to 90 days, and counteroffers of 25% to 40% are common.
How much does it cost to hire an LLM developer in India vs the US?
BLS has no LLM occupation. The closest match is Software Developers (SOC 15-1252), because retrieval pipelines, agents and fine-tuned models reach users as production software. Monthly figures are the May 2025 annual wages on the BLS profile for 15-1252 divided by 12, rounded to the nearest $10.
| Software Developers (15-1252), May 2025 |
Annual |
Monthly |
| 10th percentile |
$82,460 |
$6,870 |
| 25th percentile |
$105,210 |
$8,770 |
| Median |
$135,980 |
$11,330 |
| 75th percentile |
$171,980 |
$14,330 |
| 90th percentile |
$214,670 |
$17,890 |
US pay also moves with the city. The OEWS metropolitan area tables put the median at $213,110 a year in San Jose, $186,640 in San Francisco and $167,280 in Seattle, against $134,120 in Austin. A US employer adds benefits on top: the BLS Employer Costs for Employee Compensation release for June 2026 puts private industry compensation at $46.89 an hour, with $14.07 (30.0%) in benefits.

We publish no LLM card for India, so the India range comes from our India AI engineer rate card, the closest title: ₹250,000 to ₹970,000+ a month for engineers working from India for companies in the US, Europe and Australia.
That converts to $2,690 to $10,440+ at 92.89 rupees to the dollar (open.er-api.com mid-market rate, 22 November 2026), and it is what the engineer earns, before employer costs or a platform fee.
A developer who wires a hosted model into a support tool and one who trains adapters on your GPUs both carry the LLM title, and they sit at opposite ends of that range.
For contract work, the same card puts freelance AI engineers in India at $44 to $63 an hour at mid-level and $63 to $94 at senior level, while our LLM developer cost guide puts mid-level US freelancers at $119 to $185. Budget GPU hours and token bills on top of pay.
Which LLM skills should you screen for in India in 2026?
Our India AI engineer card lists Python, LLM APIs, RAG, vector databases, LangGraph, evaluation harnesses and MCP as the typical stack, and warns that a long list on a CV tends to mean shallow exposure. Write down whether you need someone to call hosted models, fine-tune open-weight ones or serve them, then test only the layers below that apply.
Token cost in Indian languages
Sarvam measures its tokenizer by fertility, the average number of tokens needed to represent a word, and reports that it needs fewer tokens for Indic text than other open-source tokenizers.
Fertility sets both the bill and the context budget. Give the candidate one support ticket in Hindi and its English translation, and ask for the token count and price of each on the model you run today.
Open-weight models trained in India
The Sarvam 105B model card lists 10.3 billion active parameters, top-8 routing over 128 experts and a 128K context window. The 30B sibling runs 2.4 billion active parameters and saw 16 trillion training tokens, against 12 trillion for the 105B. Ask why a model with 10.3 billion active parameters still needs GPU memory for all of its weights, and which of the two your latency budget allows.
Fine-tuning on one GPU
QLoRA fine-tuned a 65-billion-parameter model on a single 48GB GPU by training low-rank adapters on top of a frozen 4-bit model, and matched full 16-bit fine-tuning on the tasks tested. Hand the candidate a checkpoint, a labeled dataset and a baseline, and ask for an adapter plus the eval that shows it earned its GPU hours.
Serving with vLLM
The vLLM paper traced the batch-size limit in LLM serving to key-value cache memory lost to fragmentation; PagedAttention cut that waste and lifted throughput two to four times at the same latency.
vLLM 0.34.0, released on 16 November 2026, removed the old model runner and raised the default cache block size from 16 to 64. Ask the candidate to size cache memory for your peak concurrent sessions, and what they would check first after that upgrade.
Personal data under India's DPDP Act
Section 3 of India's Digital Personal Data Protection Act, 2023 reaches processing outside India when it relates to offering goods or services to people in India. A retrieval index built on Indian users' chats falls under it wherever the servers sit.
Section 17(1)(d) runs the other way: it lifts most duties when someone in India processes data on people outside India under a contract with a foreign company, though the section 8(5) duty to keep reasonable security safeguards still applies. Ask how the candidate would honor a deletion request in a vector store.
Synthetic media labels
If your product serves India, the amended IT Rules cover "synthetically generated information": realistic audio, visual or audio-visual content made or altered by a computer. Platforms must label it and, where feasible, embed provenance metadata. Ask which of your features would fall inside that definition.
How do you test an LLM developer from India before you hire?
Use a take-home of three to four hours: 200 labeled support questions in mixed Hindi and English, a document set and a baseline prompt. Plant one document in the corpus that carries an instruction aimed at the model. Ask the candidate to raise the score with retrieval, a QLoRA adapter or a better prompt, and to submit an eval report with the code.
A strong answer includes:
- A held-out split fixed before any tuning, with scores reported per language.
- Tokens per request in Hindi and in English, and cost per thousand requests for baseline and final.
- The planted instruction caught, with a test that fails if a later change lets it through.
- Error analysis that separates retrieval misses from wrong generations.
- A report a US product manager can act on without a call, written in plain English.
Follow with a 45-minute live review of the submission in English, then one scenario: a provider outage hits while your US team is asleep. Listen for how they would fall back, what they would log and how they would hand over in the morning. Our LLM engineer interview questions add material for that round.
Should you hire a contractor in India, an agency or Second Talent?
A contractor in India keeps the copyright in adapters, prompts, eval sets and serving code, under both US and Indian law, until they sign it away.
Software is not one of the nine categories of commissioned work that can be "work made for hire" under 17 U.S.C. § 101, and a transfer must be in writing and signed under § 204(a). In India, Section 19 of the Copyright Act, 1957 deems an assignment that names no period to last five years and one that names no territory to cover India only. State a worldwide territory and the full term.
|
US W-2 hire |
Contractor in India |
Agency in India |
Second Talent |
| Who employs the developer |
You |
Nobody; they invoice you |
The agency |
Second Talent; you direct the work |
| Cost basis |
Salary plus about 30% in benefits (BLS) |
$44 to $63 an hour, mid-level freelance |
Agency rate including its margin |
From $25 USD/hr, 160 hours for $3,999/mo |
| Who owns weights and code |
You, as work made for hire |
The contractor, until a signed assignment naming period and territory |
Set by the agency contract |
You; IP assignment and NDAs in the contract |
| Upfront cost |
Recruiting time or fees |
None |
Varies by agency |
None |
| If the hire fails |
You restart the search |
You restart the search |
Agency terms |
One free replacement within 90 days |
A contractor in India gives you Form W-8BEN rather than a W-9. A developer who works forty hours a week in your sprints, with no other clients and access to your training data, looks more like an employee than a vendor. Employment in India now runs under the four Labour Codes, in force since 21 November 2025, which make appointment letters mandatory.
Through Second Talent's EOR service we handle the developer's employment in India, including payroll, statutory contributions and equipment, and you direct the day-to-day work. If you have already found the person, our employer of record service covers the employment side. This is a summary, not legal advice.
How many working hours overlap with ET, CT and PT from India?
India Standard Time is UTC+5:30 with no daylight saving, so a 9:30 am to 6:30 pm day in India ends before New York starts work. We agree a later schedule before the offer, and our placements work 4 to 6 hours of daily overlap with US business hours.
While the US is on standard time, a 3:00 pm to midnight shift in India runs from 4:30 am to 1:30 pm ET: four and a half shared hours with a 9-to-5 ET team, three and a half with CT and one and a half with PT.
From 14 March 2027 the same shift gains an hour with each zone. A shift that ends at midnight needs transport and, for women employees, consent for night work under the Labour Codes, so agree it in writing.
Spend the shared hours on prompt and eval reviews and release decisions on model changes. Fine-tuning jobs and eval sweeps started before the developer signs off finish while your team works, and the results are ready for the next morning in India.
How fast can you hire an LLM developer in India with Second Talent?
The first shortlist of 6 to 8 pre-vetted profiles arrives within 24 hours of a complete brief. Our vetting covers stages 2 to 4 before you see a profile.
- Brief. You state whether the developer calls hosted models, fine-tunes open-weight models or serves them on your GPUs, plus the languages your users write in, the GPU budget and the overlap you need.
- Application review. We look for an LLM feature with users and numbers: a retrieval system with a measured hit rate, an adapter with before-and-after scores or a vLLM deployment under real load. A notebook demo with no eval does not pass.
- Skills assessment. The candidate completes the mixed-language task above, and we score the eval report along with the code.
- Live interview and references. A senior engineer reviews the work in English and pushes on serving, data handling and licenses; we confirm the notice period in the candidate's current appointment letter so your start date holds.
- Your final round. You interview the shortlist, pick one, and the developer starts once their notice period ends.
Pricing, the replacement guarantee and monthly terms are on our pricing page, and you can tell us what you need to start the search. To compare beyond India, see our LLM engineers; for wider product AI work we also staff AI developers in India, and for Indic text and speech, NLP engineers in India.