Catalogue

Models

Every model here is open-weight and served on our own hardware — nothing is proxied to a third-party API. Model IDs are the strings you pass in the model field. Prices are per million tokens unless noted; see pricing for the full table and rate limits.

Language

Chat & reasoning

Nine language models. Context is the maximum window we serve; some cards advertise more than we host stably.
Model ID	Maker	Best for	Context	Licence	In	Out
qwen3-235b-a22b	Alibaba	General, reasoning, tools	256K	Apache 2.0	$0.70	$2.40
deepseek-v3	DeepSeek	General workhorse	128K	MIT	$0.50	$1.50
deepseek-r1	DeepSeek	Reasoning, maths, hard code	128K	MIT	$0.55	$2.19
mistral-small-3.1	Mistral AI	Fast multimodal (vision)	128K	Apache 2.0	$0.10	$0.30
devstral-small	Mistral AI	Coding agents, SWE tasks	128K	Apache 2.0	$0.10	$0.30
gemma-3-27b	Google	Efficient multimodal	128K	Gemma Terms	$0.12	$0.30
phi-4	Microsoft	Small, strong reasoning	16K	MIT	$0.07	$0.14

Audio

Speech to text, text to speech

Both audio models are billed on their natural unit rather than tokens: transcription per minute of input, synthesis per million characters of output.

whisper-large-v3

MIT

OpenAI · speech-to-text

Transcription and X→English translation across 99 languages. Returns segments or word-level timestamps; handles noisy, accented, and code-switched audio well.

Endpoint/audio/transcriptionsLanguages99 Timestampsword / segmentPrice$0.004 / min

kokoro-82m

Apache 2.0

Hexgrad · text-to-speech

Lightweight, natural 24 kHz synthesis with 54 voices across 8 languages. Low latency to first audio; streams PCM or returns WAV/MP3/Opus.

Endpoint/audio/speechVoices54 Sample rate24 kHzPrice$0.80 / 1M ch

Retrieval

Embeddings & reranking

For RAG pipelines. Embeddings are priced per million input tokens; reranking per million tokens scored.
Model ID	Maker	Task	Dim / context	Licence	Price
bge-m3	BAAI	Embeddings (multilingual)	1024 / 8K	MIT	$0.02 / 1M
bge-reranker-v2-m3	BAAI	Cross-encoder rerank	— / 8K	Apache 2.0	$0.05 / 1M

Model IDs are stable. When we retire a snapshot we keep the ID pointed at the newest compatible weights and give 30 days’ notice by email and on the status page. Pin a dated snapshot (e.g. deepseek-v3-0324) if you need bit-for-bit reproducibility.