Meet us at Singapore FinTech Festival, 18-20 Nov, Hall 6, booth 6F07-08. Book a time →

Customers / TransVirtual

TransVirtual

Code review on every pull request.

TransVirtual, an Australian transport and logistics software company, reviews every pull request into main on Xinference Model API.

~10 secondsper pull request review
Every pushreviewed on pull requests into main
Requireda posted review is a merge check
One settingto switch the model
How we worked these out

Cost compares list prices per million tokens. DeepSeek V4.1 Flash is $0.30 input and $1.20 output, Claude Sonnet 5.5 is $2 input and $10 output, so 85% less on input and 88% less on output.

Output speed is from Xinference testing of DeepSeek V4.1 Flash at 10K input tokens using the Artificial Analysis method, against 205 tokens/s for DeepSeek’s own API.

Review time is as reported by TransVirtual.

Model prices are list prices in US dollars, as of November 2026.

Company
TransVirtual
Industry
Transport and logistics software
Product
Model API
Model
DeepSeek V4.1 Flash, thinking off
Workload
Code review in CI

The challenge

TransVirtual reviews every pull request into main each time a developer pushes. Those reviews ran on a closed-model API. To run on every push, a review has to come back in seconds, not minutes.

What they built

A review bot now runs on Xinference Model API. Each review posts comments on the exact lines plus a summary, and a posted review is a required check before merging.

  1. 1A developer pushes to a pull request into main.
  2. 2The change goes to Xinference Model API for review.
  3. 3The bot comments on the exact lines and adds a summary.
  4. 4The posted review becomes a required check before merging.
POST /v1/chat/completions
client = OpenAI(
    base_url=XINFERENCE_BASE_URL,
    api_key="XINFERENCE_API_KEY",
)

The result

With DeepSeek V4.1 Flash and thinking off, TransVirtual reports about 10 seconds per review, fast enough to run on every push. The team no longer depends on the closed model for reviews, and can switch models with a settings change.

Before and after

Cost

BeforeClaude APIclosed model, priced per token
NowDeepSeek V4.1 Flashon Xinference Model API

Up to 88% less cost

per token than Claude Sonnet 5.5, at list prices as of November 2026

Speed

DeepSeek’s own API205 tokens/soutput speed
Xinference Model API~240 tokens/soutput speed

About 16% faster

output than DeepSeek’s own API, on DeepSeek V4.1 Flash.

We needed a review that keeps up with every push. With DeepSeek V4.1 Flash and thinking turned off, it's back in about ten seconds, so we made it a required check on main. And we're not tied to a closed model anymore.
Senior Software Engineer, TransVirtual

More customer stories

ShiftCare

Support agents for NDIS providers, on inference hosted in Australia.

AgentsDatapel

Every cycle count checked by an agent, signed off by a person.

AgentsDetrack

Every new trial gets a sales brief in minutes.

Agents

Try it on your own workflow.

Point an OpenAI-compatible client at Model API and pick a model.