A regional Azure option, with separate indexes and honest language limits.
What changed for the pilot
We successfully deployed and invoked text-embedding-3-large as Regional Standard in South Africa North. It becomes an optional beta through the same Embiro embeddings API. Qwen remains available; neither model is promoted as validated for Luganda.
The Azure option uses 3072 dimensions, compared with Qwen’s 1024. Switching requires a separate model-specific index and re-embedding both documents and queries. Matching dimensions alone would not make different embedding spaces interchangeable. No automatic fallback between these models is configured.
A useful Swahili signal, not a language leaderboard
Both endpoints placed all four English answers first. Azure placed three of four Swahili answers first and all four in the top three; Qwen placed two first and two in the top three. Four queries are too few for a reliable population-level ranking.
Both placed just one of four Luganda answers first. Qwen recovered two in the top three, Azure one. The next quality gate is a larger native-reviewed set, including local domain vocabulary and relevant documents written in the target languages.
Read the concurrent timings with the throttles
Azure’s median was 108.84 ms for sequential requests and 108.16 ms at concurrency four. Qwen’s medians were 140.18 ms and 143.79 ms. These are deployed endpoint timings from one client, not pure compute speed.
The Qwen endpoint returned six HTTP 429 responses during the eight-request concurrent setting. Bounded ten-second waits made the longest successful request 30.38 seconds. Azure returned no 429 responses in this sample. Our production gateway’s admission controls differ from this direct-endpoint probe, so this is not a production user-latency prediction.
An initial client setup failed certificate verification because only the private GPU CA was loaded. We corrected the client to extend the system trust store. TLS verification remained enabled; that failed setup attempt is excluded from measured samples.
Pricing and processing boundaries
The introductory Embiro customer rates are $0.25 per million input tokens for the Azure option and $0.01 for Qwen. For 100,000 measured input tokens this is $0.025 and $0.001 before per-request micro-dollar rounding. These are customer prices, not measured provider costs or GPU break-even estimates.
The successful measured evaluation consumed 772 Azure input tokens and 816 Qwen input tokens. We did not resolve an exact South Africa provider retail meter, so no upstream regional unit-cost claim is made. Shared GPU idle time, networking, storage and taxes are not assigned to these requests.
Project location alone does not guarantee processing location for Global model deployments. We specifically deployed Regional Standard; the applicable boundary is the Azure geography. This does not certify every supporting service, abuse-monitoring process or console request path as Africa-only.
Results, in context.
| Language | Queries | Azure 3 Large | Qwen 0.6B |
|---|---|---|---|
| English | 4 | 4 / 4 | 4 / 4 |
| Swahili | 4 | 3 / 4 | 2 / 4 |
| Luganda | 4 | 1 / 4 | 1 / 4 |
Four questions per language against the same 16 English candidate documents. Not a representative language benchmark.
| Endpoint | Concurrency | Samples | Median | Maximum | 429 responses |
|---|---|---|---|---|---|
| Azure 3 Large | 1 | 8 | 108.84 ms | 238.91 ms | 0 |
| Azure 3 Large | 4 | 8 | 108.16 ms | 130.04 ms | 0 |
| Qwen 0.6B | 1 | 8 | 140.18 ms | 141.00 ms | 0 |
| Qwen 0.6B | 4 | 8 | 143.79 ms | 30381.56 ms | 6 |
End-to-end direct HTTPS request timings include response parsing and retry waits. Maximum equals nearest-rank p95 for eight samples; requests eventually succeeded.
How we ran it.
- One paired retrieval pass: 16 English support documents, four English queries, four Swahili queries and four Luganda queries. Each has one authored relevant answer; translations have not received native-speaker review.
- Each endpoint embeds the same documents and queries. Cosine similarity ranks full, unnormalised returned vectors; similarity itself normalises by vector lengths. No vectors are mixed between models.
- Azure text-embedding-3-large version 1, Regional Standard in South Africa North, returns 3072 dimensions. The existing Qwen3 Embedding 0.6B Cape Town service returns 1024 dimensions. Azure accelerator, inference runtime and precision are undisclosed. Qwen uses fp16 revision 97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3 on our existing shared NVIDIA T4 service; this is endpoint-level testing, not matched hardware.
- Client: Python 3.12 urllib on an Azure Standard_B2s_v2 in South Africa North. Retrieval calls precede timings. Eight short single-text requests at concurrency 1 and eight at concurrency 4 per endpoint; no cold-start study or throughput-saturation sweep. Timings run from HTTPS dispatch through response parsing, including retry waits.
- No customer data, licensed dataset recordings or existing search indexes were used or modified. Fixtures are synthetic.
The limits matter.
- Only four queries per language, authored synthetic labels and unreviewed translations; no statistical significance or broad multilingual quality ranking.
- One retrieval pass and eight latency samples per setting; nearest-rank p95 equals the maximum. No long-document, adversarial, sustained-load or outage evaluation.
- Different models, vector sizes, serving software, hosting regions and shared-load conditions. Managed runtime/precision are undisclosed; no isolated hardware comparison.
- Concurrent Qwen latency includes six throttles and bounded retry waits. Production admission, queueing, authentication and frontend overhead are outside these direct-endpoint timings.
- No native-speaker Luganda validation and no guarantee of African-only ancillary data handling. Customer pricing is distinct from actual provider cost.
Evidence & sources.
The public extract contains the methods, selected results and caveats in this note. It is a curated summary, not the full raw run. Fingerprints identify the retained source artifacts; gated audio and private operational records are not included.
Download results extract- Azure deployment types and processing geography
- Azure text-embedding-3-large model card
- Qwen3 Embedding 0.6B model card
Source artifact fingerprints · SHA-256
a6d3a9cbf07dedff1157bbb4e901033f74579d26f4c242430f57a6848b21489aa9cc68dc8d261c2ca12cea203d55411bcea9567a5a87fd93148ef012adb8e2fcfa53c7755082ff59a5e9b07410a00adbe3d17f9c1939e4790842acb3978fd5d1af19839dcc96d543a4ee31272779704e6cec85763b1fdc21c6b083d025139bbd23df17428d0391d73d524be1d89659bf0b88b9f522237f9a4dfd53657ad8c6674f9f13fb6649af407567c520b69709fab3d65d1557320302af4b6317a5b62ebc416067653049c05ab587072837c2d606e188f0957663cfc7619e9bb76a6d4eb1Suggested citation: Embiro Labs Research. “Regional embeddings: a small Swahili retrieval check” Published 20 Oct 2026. embiro.ai/research/foundry-regional-embeddings
Report a correction
