Hellas Inference
Evidence

Measured, not claimed.

Every number below was measured by us on our own GPUs; derived estimates are marked ≈ and footnoted.

  • / GPU
    concurrent summaries
  • / 5c
    RAG quality
  • −%
    latency cut
  • TTS streams
  • / H200
    voice calls
  • ≈–×d
    cheaper than frontier APIs
The stack we ship
  • NVIDIA H200
  • vLLM FP8
  • prefix caching
  • Iris 8B
  • Metis 12B · 26B
  • Clio
  • Calliope
Benchmark record

What we have measured

01

Greek IVR summarisation

Summaries of real Greek telecom contact-centre calls. Iris 8B (built on Krikri-8B), FP8, vLLM, one H200.

200
concurrent per GPU within a 2-second KPI
−45%
latency at 200 concurrent, by placing instructions before the transcript
≈ 1/30a
cost per call versus a frontier API
IVR summary latency at 200 concurrent requests, before and after prompt reorderingRelative latency falls from 100% to 55%, a 45% reduction, with the same summary quality.Before100%After55%
Relative latency at 200 concurrent; instructions moved before the transcript.

Why it matters: summarisation moves off US APIs at a fraction of the cost, with quality parity verified.

02

Greek RAG helpdesk (Pythia benchmark)

Answers over Greek enterprise documents, judge-scored on a 5-point scale. Dense retrieval, whole-document context, open model.

4.41c
Gemma-4-12B, level with the best frontier models
4.18
Krikri-8B
+0.20
faithfulness from OCR enrichment of screenshots
Judge-scored answer quality on Greek enterprise documentsGemma-4-12B scores 4.41 of 5, equal to the frontier ceiling of 4.41; Krikri-8B scores 4.18.012345Gemma-4-12B4.41Krikri-8B4.18Frontier ceiling4.41
Scale 0–5; the dashed outline is the best frontier model we tested.c

Why it matters: open models match frontier quality on Greek documents. Dense retrieval beat hybrid BM25 + dense on this corpus, so we keep the simpler pipeline.

03

Agent model: tool calling

Booking and service tasks in our voice-agent test harness.

131 / 190
tasks completed correctly, Gemma-4-26B-A4B
8 / 190
tasks completed correctly, Krikri-8B
Booking and service tasks completed correctly, out of 190Gemma-4-26B-A4B completes 131 of 190 tasks; Krikri-8B completes 8 of 190.Gemma-4-26B-A4B131 / 190Krikri-8B8 / 190
The faint track is the full set of 190 tasks.

Why it matters: the right model for each role. Gemma 4 drives the tools; Krikri stays the model for Greek contact-centre text.

04

Greek text-to-speech

Orpheus-3B, the base of Calliope, on vLLM, one H200. Measured before the voice adapter was merged; a merged adapter does not change the size of the model.

423
concurrent streams per GPU
423 concurrent Greek TTS streams on one H200A grid of 423 dots, one per concurrent stream.
Each dot is one concurrent Greek TTS stream on a single H200.

Why it matters: a voice minute costs ≈ under €0.01 in TTS compute.b

05

Voice, end to end

Clio speech recognition, voice-activity detection, the Metis agent model and Calliope voice on one H200.

8
concurrent calls per H200 with every caller turn heard; 16 with four speech-recognition workers
4.55 / 6.66 s
p50 / p90 from the end of the caller's speech to the first answer audio, at 8 calls
1.92 / 2.53 s
p50 / p90 with one call on the GPU
32 / GPU
speech-recognition streams with batching
End-to-end voice response time at 8 concurrent calls per H200p50 is 4.55 seconds and p90 is 6.66 seconds, against a 2.0-second KPI.0 s2 s4 s6 sp504.55 sp906.66 s2.0 s KPI
The dashed line marks the 2.0 s KPI. This is the number we are working hardest to lower.b

Why it matters: this is our sizing basis for large contact-centre deployments, and the number we are working hardest to raise.

06

Serving know-how

Quantisation and scheduler settings on Hopper GPUs.

8×
slower: AWQ INT4 versus FP8 on H200
256
optimal max-num-seqs in vLLM
Relative throughput on H200: FP8 versus AWQ INT4FP8 delivers 8 times the throughput of AWQ INT4 on H200.FP88×AWQ INT41×
Relative throughput on H200, higher is better.

Why it matters: operational details decide the cost per call. We measure them rather than assume them.

  1. a ≈1/30 cost vs frontier API: estimate from measured H200 throughput × rental cost per hour against public list prices of a frontier API for the same token volume, Nov 2026.
  2. b Voice minute under €0.01: TTS compute only; end-to-end voice (ASR + LLM + TTS) measured at 8 concurrent calls per H200 with p50 4.55 s / p90 6.66 s response time, callers paced at phone speed, telephone-network delay not included (26 October 2026).
  3. d ≈10–250× cheaper than frontier APIs: estimated range across our workloads, from measured H200 throughput × rental cost per hour against public list prices of frontier APIs for the same token volume, Nov 2026. The ratio depends on the task, the prompt length, GPU utilisation and which frontier model is compared; Greek call summarisation sits at ≈30× (note a).
  4. c Quality scores: frontier model as judge with a fixed rubric, Greek reference answers; judge and rubric published in the full record.
Measured vs claimed

We publish the stack, the settings and the data cleared for evaluation. If a number here cannot be reproduced on the stack we ship, we take it down.

Method

How we measure

  • The same stack we deploy: vLLM, FP8, NVIDIA H200.
  • Quality is scored by frontier models acting only as judges, on data cleared for evaluation.
  • Paired comparisons: the same data, one change at a time.
  • Benchmarks on rented GPUs use non-customer data only.

Model roles. Iris: the instant model, for Greek contact-centre text and call summaries. Metis: the thinking model, for document Q&A and agents that call tools. Clio: speech to text. Calliope: voice. Each is an open-weight base model with our own fine-tuned adapter; the base is named on every card. Llama and Qwen stay available as routing alternatives.

Next

A public business-truth benchmark

How often an answer is true to the business, with and without the Company Brain, on Greek data, comparing small open models with frontier models.

Every shipped fine-tuned model already carries an AI Act (Regulation (EU) 2024/1689) dossier and a model card with training-data manifests.

Our research programme

Measure it on your data.

We run the same benchmark on a sample of your calls or documents before you commit.

Request a demo[email protected]