The specialty inference house for the open-weights models hyperscalers under-serve.
OpenAI-compatible API. Embeddings, rerankers, small language models, and African-language frontiers — hosted where you need them. EU-resident at launch; Lagos & Nairobi H2 2026.
1# OpenAI-compatible. Drop-in replacement. 2curl https://api.tomoul.ai/v1/chat/completions \ 3 -H "Authorization: Bearer $TOMOUL_KEY" \ 4 -H "Content-Type: application/json" \ 5 -d '{ 6 "model": "tomoul/inkubalm-0.4b", 7 "messages": [ 8 { "role": "user", 9 "content": "Habari za asubuhi" } 10 ] 11 }'
Models other clouds either run badly, or don't run at all.
Embeddings and rerankers because that's where the margin is. Small models because they're sharp and cheap. African-language frontiers because nobody else hosts them well. Every model is OpenAI-compatible at the API layer.
Multilingual, multi-granularity embedding. Dense, sparse, and ColBERT in one pass — the workhorse for retrieval over 100+ languages.
Instruction-tuned dense embedding. Best-in-class on MTEB at the <2B-param tier. Drop it behind your search and watch recall rise.
Cross-lingual encoder fine-tuned on 17 African languages. The reference model for Swahili, Hausa, Yoruba, Amharic classification.
14B reasoning-tuned. Frontier-class reasoning at sub-frontier cost. Tool use, structured output, and 16K context.
Apache-licensed, instruction-tuned 20B. The best general-purpose chat model you can self-host — we just host it for you.
A tiny, regionally-tuned LLM for Swahili, IsiZulu, Yoruba, Hausa, IsiXhosa. Lands with the Lagos POP H2 2026.
Run models where you need them.
No-train by default. EU-resident at launch. Africa POPs land H2 2026 — Lagos first, then Nairobi. The same OpenAI-compatible API everywhere.
Pay your way.
Stripe and USDC at launch. Local rails — Flutterwave, Paystack — roll out with the Africa POPs. No procurement officer left behind.
Card, Apple Pay, Google Pay, SEPA. Auto-top-up + invoicing.
On-chain credit top-up via Circle. Best for crypto-native teams.
Local mobile-money and card rails for Kenya, Nigeria, Ghana, South Africa, Egypt. Ships with Lagos POP.
One engine, three surfaces.
Cloud is the product. CLI is the on-ramp. Engine is the open-source proof. They share a model catalog, an API shape, and a Zig core.
Inference API, hosted
OpenAI-compatible endpoints for chat, embeddings, rerankers, transcription. Per-token pricing. EU at launch.
Browse the catalog →Local + cloud
A single static Zig binary. tomoul serve bge-m3 runs locally. --cloud points at api.tomoul.ai.
Open Zig core
12MB binary. CUDA, Metal, CPU. @import("tomoul") — and you're inferencing.
The API your stack already speaks. The models the others won't host.
Sign up. Get a key. Stream your first token in 60 seconds.