API KEYS2 ACTIVE
Productionsk-prod-••••••••4Q7K
ACTIVEStagingsk-stage-••••••••8N2P
ACTIVELAST ROTATED · 02 AUG 2026ROTATE KEY ↗
Outcome-first AI inference
for enterprises.
Better model choices.
Nothing else has to change.
We help you decide which model to use for each job, so you save cost without ripping anything out.
→QUERY STREAMLIVE
480M TOKENS / MONTHtraffic detected
DVAR MODELS74%
FRONTIER26%
ILLUSTRATIVE MODEL COSTLOWER IS BETTER
OPEN API INDEX100
BLENDED INDEX31
−69%lower cost
MODEL ALLOCATION
Routine knowledge work stays on Dvar; complex analysis can still reach a frontier model.
Evals, routing, and traffic in one place.
The portal shows the eval pass rate, the model serving each request, and the bill. Connect the endpoint, verify the serving path, and reach your Dvar engineer directly.
APP.DVARLABS.AI / INFERENCE LIVE OPERATIONS
Production APIapi.dvarlabs.ai · India region · OpenAI compatible
ALL SYSTEMS HEALTHYREQUESTS / MIN2,148+12% / LAST 24H
P95 LATENCY180 msSTEADY
UPTIME99.98%30 DAY WINDOW
HEALTH CHECKS30S POLL
HEALTHY99.98% UPTIME
Endpoint180 ms
Model loadREADY
Policy layerENFORCED
LIVE REQUEST TRACELAST 60 SEC
Proof on tasks like yours.
Measured economics.
A live comparison across three bank-adjacent tasks, judged against a written answer key. Dvar Balanced matched the required outcomes at 12×–61× lower cost in these single-shot runs.
LIVE PRODUCT COMPARISONDvar Balanced vs Claude Opus 5
18 AUG 2026 · INRDOCUMENT EXTRACT
Every field, at a lower unit cost.
Both returned all 10 required fields against the answer key.
LOWER DVAR COST12×
MEASUREDVAR BALANCEDCLAUDE OPUS 5
Cost / run
₹0.12
₹1.53
Time9.3s5.2s
Output tokens160186
Live product comparison, 18 Aug 2026. Same prompt, new conversation each run, and quality judged against a written answer key. Single-shot results, not an SLA; card and onboarding examples are synthetic.
Evals first. Then a better choice.
Then a lower bill, if you want it.
You are not changing the stack. We test models on your work, help you choose better from what you already have, and when you want to save more, give you our open-source models. Same quality, lower price, work stays in India.
01Evals
See what actually works.
We test models on your real work, not a public leaderboard. You see quality, speed, and cost, so the next choice is obvious.
See how we work02The right model
Better choices. Same setup.
Those tests tell you which model does the job from what you already have. You keep the apps, the workflow, and the API. You just stop overpaying.
See the path03Open-source models
Same quality. Lower cost. Sovereign.
When you want to save more, you can use our open-source models. Same quality, lower price, work stays in India, still without changing how your team builds.
See the modelsFine-tuning for the work that deserves it.
The eval decides when fine-tuning is worth it. When it is, Dvar engineers co-build a smaller, stronger model for the work that repeats: curate examples, tune an adapter, evaluate on held-out tasks, and run the improvement loop with your team.
Dvar model service live checkpoint
Collect examplesCustomer cases and reviewer labels
doneTrain adapter12B checkpoint · three passes
currentEvaluate releaseHeld-out cases before serving
nextNext review18 Aug · held-out set→
TALK TO AN ENGINEER
Everyone sells an API.
Dvar sends an engineer.
Bring one real workload. We build the eval with your team from your own traffic, benchmark the models against it, and stay on the account from pilot through production.
Book a workload review