Context before conclusions
Hardware, precision, sample size and timing boundaries accompany every study. A quick response is only useful if the answer is right.
EMBIRO LABS / RESEARCHLess guesswork.
Less guesswork.
More evidence.
What it takes to make AI useful, affordable and closer to home. Notes from our experiments in African inference.
Explore the research15 research notesMethods includedFailures included, too
Includes listening comparisonsFEATURED EXPERIMENT / SOUTH AFRICA
Does every AI workload need a GPU?
Kokoro, embeddings and three transcription models on Azure CPUs in South Africa, compared with our AWS Cape Town baseline.
Kokoro is a promising CPU candidate.Interactive transcription still favours the GPU. The details—and the tradeoffs—matter.
Read the findings THE NOTEBOOK
Every test tells us something.
Our early experiments, practical decisions and questions we’re still working on.
15 research notesNewest studies first · Dates reflect the experiment
Infrastructure
FIELD NOTE / 15
Does every AI workload need a GPU?
Kokoro, embeddings and three transcription models on Azure CPUs in South Africa, compared with our AWS Cape Town baseline.
1.44sKokoro EN median on Azure D4
Speech
FIELD NOTE / 14
Faster speech, one language at a time.
Batching shortened one long English Orpheus run by 21%. Luganda and Swahili told a different story.
21.1%less generation time · one English trial
Speech
FIELD NOTE / 13
Kinyarwanda: promising is not ready.
Forty recordings show why language coverage on a model card needs a local evaluation before a pilot release.
35.83%word error rate · 40 recordings
Speech
FIELD NOTE / 12
Choosing speech models for local languages.
Sunflower, Crane and Whisper for transcription; Orpheus, Spark and Kokoro for generation. The fastest option did not always sound best.
12.5%optimised Sunflower Swahili WER
Speech
FIELD NOTE / 11
Bonjour, Kokoro.
A small French voice feasibility test, followed by broader CPU comparisons.
3French prompts tested
Speech
FIELD NOTE / 10
Testing Crane on Luganda and Swahili.
A matched recording comparison shows a meaningful improvement over Whisper, and why Luganda still needs care.
24.3%Crane Swahili WER · 12 clips
Infrastructure
FIELD NOTE / 09
How much media fits on one T4?
Speech, images and a small video experiment reveal where a compact shared worker starts to strain.
9failed shared SDXL requests retained
Serving
FIELD NOTE / 08
From a voice sample to a usable API.
Local Kokoro integration checks covered queued jobs, idempotency and accounting before cloud speech rollout.
2local execution environments verified
Serving
FIELD NOTE / 07
More concurrency is not more capacity.
Gemma, embeddings and reranking behind a real gateway: what changed between two, four and eight concurrent clients.
183 / 183final measured requests completed
Models
FIELD NOTE / 06
Seven small models. No universal winner.
Gemma, Qwen, Llama, GLM, Ministral and Kimi on one T4, with language and reasoning tradeoffs exposed.
7serving profiles screened
Models
FIELD NOTE / 05
Bigger did not always answer better.
A non-thinking Qwen3.5 comparison against Qwen3-8B and Gemma 3 4B, including a separate sampling check.
182measured requests in the main run
Serving
FIELD NOTE / 04
Silence should stay silent.
A Whisper adapter retry caught a decoder incompatibility and a hallucinated transcript on silent audio.
0.679smedian · six warm synthetic clips
Models
FIELD NOTE / 03
An API response is not an answer.
Our first capability screen separated transport success, retrieval quality, vision and incomplete media tests.
24warm chat responses · three models
Infrastructure
FIELD NOTE / 02
Warm GPUs, cold reality.
Our first Cape Town CPU-versus-T4 comparison explains why readiness needs a real inference probe.
51.03 → 1.39sQwen long-prompt first token · CPU to T4
Serving
FIELD NOTE / 01
Our first benchmark: fast tokens, unfinished answers.
A local DeepSeek test established the gateway and accounting baseline before cloud comparisons began.
10 / 10short runs hit the output cap
HOW TO READ OUR RESEARCHEvidence with
Evidence with
the edges left in.
These are first-party engineering notes from an early pilot, not peer-reviewed papers or performance guarantees.
Keep the awkward results
Failed runs, small datasets and unresolved quality questions stay in the record. Historical experiments do not imply current model availability.
Show our working
Each note includes a downloadable results extract and source fingerprints. Gated recordings, credentials and operational identifiers stay private.
Follow the experiments.
New notes join this notebook as benchmark results are reviewed and published.