Step 1
Train
Qwen3.8 27B, post-trained with RL on your workflow data. It learns your job, not the internet's.
Trainingrun E1090 · verified
Mean NLL by step · qwen3.8-27b · LoRA r16 · 3 epochs
open models, trained on your work
The model that knows your work.
Replace your expensive AI API with an open model trained on your work. Proven before it ships.
What changes:
- 1 /One call: no chain of calls, no retry loop.
- 2 /Knows your business: your terms, customers, language and locale are in the weights, not pasted into every prompt.
- 3 /Faster: fewer calls on a smaller model.
how it works
Train it. Own it. Run it.
- 1 /Train. We write your quality bar down as a benchmark, from your spec and your best human examples. Then a training loop searches recipes until an open model passes it in a single model call.
- 2 /Own. You keep the model, the benchmark, and the record of what went into training.
- 3 /Run. The new model runs in shadow behind your existing endpoint, then takes 5%, 25%, 100% of traffic. Your client code stays the same, and one step rolls it back. Or you run it yourself.
custom models
Test models on your work. Build one when the evidence supports it.
We test candidate models on your workflow, then train and serve a custom model when the results justify it.
Step 2
Prove parity
Held-out evals against GPT-5.6 Sol on your real tasks. It ships only when the results support it.
Evalsbeats GPT-5.6 Sol
Held-out eval · 200 rows / arm · 95% CI
Step 3
Run 10× cheaper
Served on our managed stack at $0.40/$3 per 1M tokens. Sol is $5/$30.
Inferenceserving · 78% traffic
TTFT p50
140ms
Decode
75tok/s
KV-cache hit
94.9%
vs Sol tokens
10×cheaper
qwen3.8-27b · managed serving · last 24h
Step 4
Switch without the leap
Shadow traffic proves the model is operationally sound. Endpoint A/B tests the result with the same URL and keys. No client release is required to ramp or roll back.
RolloutA/B live · 5% variant
live traffic
one endpoint95%control
gpt-5.6-sol5%variant
qwen3.8-27b · tunedSame endpoint, API, and keys · shadow mirrors traffic with responses discarded
case 1
Case 1 · Bithumb (crypto exchange, Korea)
The customer ran an agentic workflow on a frontier API to write research reports. We trained an open model to write the same reports, from the same evidence, behind the same release checks, in one call.1
| 30 days | tokens | |
|---|---|---|
| frontier workflow | $130,328 | 14.790B |
| trained model (est.) | $4,604 | 4.843B |
| change (estimated) | -96.5% | |
The trained model is running in shadow. Cost only; latency not compared.2
It passed 402 of 402 QA checks and scored 0.990 on semantic review, where a judge model checks each claim against the evidence (frontier workflow: 0.985).
read the full case 1 →what you keep
What you keep
- the trained model and adapters
- the benchmark, as code
- the data record
- the shadow ledger
- the right to leave with all of it
next