Our own models.
Language models for Romanian and English, speech for agents, a decision model that sees, and the open-source engine that trains and serves them.
Surogate 3.7 Flashllm.buzz Open weights, text and images in, served from the EU on llm.buzz.
- tokens of context
- 1M
- tokens of output
- 32K
- per million input tokens, Surogate Spark EU
- $0.15
Surogate 3.5
Romanian-first assistants with thinking in Romanian and English. Apache 2.0.
- Surogate 3.5 9B57.27 chrF2, English to Romanian translation
- Surogate 3.5 4B53.61 chrF2, English to Romanian translation
- Surogate 3.5 2B47.42 chrF2, English to Romanian translation
Measured by us, as each model card reports it.
Decisions, speech and the engine.
- Decision modelSurogate RuneText or an image in, one of your options out, with a calibrated probability for every one.
- parameters
- 26B
- tokens of context
- 262k
- median for a decision about an image
- 0.18 s
- Speech for agentsSurogate SpeechSmall open speech models that run on a CPU. Jackrabbit listens, Amami speaks.
- Amami TTS
- 2.83% WER
- Jackrabbit ASR
- 5.69% WER
- Jackrabbit Streaming ASR
- 0.72 s to text
- Training and serving, open sourceSurogate EngineTraining and serving in one toolkit, on native C++/CUDA engines. Fine-tune, run RL, then serve the result over familiar HTTP APIs.
- training vs. Unsloth
- 2.53×
- serving vs. llama.cpp
- 7.0×
- licence
- Apache 2.0
Use them on llm.buzz.
Surogate 3.7 Flash is on llm.buzz, with one key for every model.