KV-cache transfer engine that prefills small models and decodes large ones.
OneTriangle inference maps a small model key-value cache into a larger model so prefill runs on the small model and decode stays on the large one. Eligible pairs include Llama and Qwen size jumps, and ordinary prefill takes over when a pair misses quality or latency gates.
Customers call an OpenAI-compatible API for managed open-weight models including Llama, Qwen, Mistral, Gemma, and DeepSeek. The company publishes pair-level transfer research and is upstreaming the method into vLLM.
Tracer
Echo
tracerml.ai
Fireworks AI
Fireworks AI Inference
fireworks.ai
Groq
GroqCloud
groq.com
Baseten
Pre-optimized Model APIs
baseten.co
Together AI
Together AI Inference
together.ai
| Attribute | OneTriangle — KV-cache inference | Echo | Fireworks AI Inference | GroqCloud | Pre-optimized Model APIs | Together AI Inference | |
|---|---|---|---|---|---|---|---|
| Provider |  | | | | | | |
| Founded | 2026 | 2026 | 2022 | 2016 | 2019 | 2022 | |
| Sells To | Developer | Enterprise | Enterprise | Enterprise | Enterprise | B2B, Enterprise | |
| Pricing Model | Usage-based | B2B Subscription | Recurring | Recurring | Recurring, Software, Usage-based | Recurring, Subscription, Usage-based | |
| Ownership | Venture-backed | Private | Venture-backed | Venture-backed | Venture-backed | Venture-backed |
Provider
Founded 2026
Sells To Developer
Pricing Model Usage-based
Ownership Venture-backed
Provider
Founded 2026
Sells To Enterprise
Pricing Model B2B Subscription
Ownership Private
Provider
Founded 2022
Sells To Enterprise
Pricing Model Recurring
Ownership Venture-backed
Provider
Founded 2016
Sells To Enterprise
Pricing Model Recurring
Ownership Venture-backed
Provider
Founded 2019
Sells To Enterprise
Pricing Model Recurring, Software, Usage-based
Ownership Venture-backed
Managed service for fine-tuning open-source models on dedicated or shared GPU infrastructure for production use.
Baseten Inference Stack for high performance with custom kernels, decoding, caching on optimized infra.
Weave Router is a drop-in proxy that routes each inference request to the lowest-cost model that answers at frontier parity.
Cloudflare AI: serverless inference, agents, RAG, and secure runtime primitives on the global edge.
Async inference API for long-horizon AI agents, serving leading open-source models through OpenAI- and Anthropic-compatible endpoints.
An open-source AI gateway and proxy that calls many large language model APIs in the OpenAI format.
Crusoe Managed Inference deploys and scales production LLM workloads with breakthrough speed and throughput.
InferenceX continuously benchmarks LLM inference throughput, latency, and cost across GPU vendors.
Fastest cloud for training/inference of 1B-24T models with no sharding.
IonRouter is a high-throughput, low-cost inference API and drop-in OpenRouter alternative powered by Cumulus Labs' IonAttention engine.
Vast.ai Serverless runs autoscaling GPU inference endpoints on managed capacity.
Managed service for fine-tuning open-source models on dedicated or shared GPU infrastructure for production use.
Baseten Inference Stack for high performance with custom kernels, decoding, caching on optimized infra.
Weave Router is a drop-in proxy that routes each inference request to the lowest-cost model that answers at frontier parity.
Cloudflare AI: serverless inference, agents, RAG, and secure runtime primitives on the global edge.
Async inference API for long-horizon AI agents, serving leading open-source models through OpenAI- and Anthropic-compatible endpoints.
An open-source AI gateway and proxy that calls many large language model APIs in the OpenAI format.
Crusoe Managed Inference deploys and scales production LLM workloads with breakthrough speed and throughput.
InferenceX continuously benchmarks LLM inference throughput, latency, and cost across GPU vendors.
Fastest cloud for training/inference of 1B-24T models with no sharding.
IonRouter is a high-throughput, low-cost inference API and drop-in OpenRouter alternative powered by Cumulus Labs' IonAttention engine.