Ramp launches its own AI model router
Two moves today point at the same prize: whoever sits between you and the model gets paid. Ramp is building the toll road, while Liquid AI is making the smaller models running on it go faster.
Expense-management fintech Ramp just opened its own AI model router, letting developers and companies route prompts across OpenAI, Anthropic, DeepSeek, and more through one API. The service, called Router, launched Wednesday evening and is available only in the United States; it is free to use through the end of 2026 (users still pay underlying inference costs) and ships with a $26 launch credit. Ramp says it has quietly run this router internally for its own AI workloads for three years, and the product looks a lot like OpenRouter, though OpenRouter still exposes far more model choices. Where it gets interesting is the routing logic: customers can steer traffic by provider flex tiers, let Router pick a model based on up to three chosen benchmarks, or send only the hard problems to expensive models.
For Ramp, this is a two-pronged bet. It taps the booming inference-services market and slots neatly into the company's existing token-spend tracking and cost-management tools — a natural wedge to sell expense software to AI-heavy teams. It also follows Stripe's recent sprint into the same "toll house for AI inference" space, suggesting the routers that sit between apps and models are becoming their own contested layer. One caveat worth flagging: Router keeps model inputs, outputs, and tool calls for a year by default, pledging to strip personally identifiable information before using any of it to improve the product — a retention stance buyers should read before they pipe sensitive data through it.
Liquid AI shipped DSpark draft models for its LFM2.5 family that speed up inference up to 3.2x on a GPU and nearly 3x on a MacBook, with no change to output quality. DSpark is a speculative-decoding technique: a small draft model proposes several tokens at once, the target model verifies them in a single pass, and because greedy output must match the target's distribution, the emitted text is mathematically identical to running the big model alone. The practical payoff lands hardest on the edge — Liquid AI reports function-calling latency drops 57% on average for the 2.6-billion-parameter model, which matters for agents that fire many tool calls in a row. The draft models are open weights, with day-one support upstreamed into both llama.cpp and SGLang, so the speedup is reachable on a laptop, not just a datacenter H100.
This is the same family of acceleration we looked at with DFlash 2's parallel drafting — but Liquid AI's release is notable for how broadly it is packaged: three model sizes, both major local inference engines, and a claim that quality is preserved by construction rather than approximated. For teams running small open models on-device, that combination pushes agentic workloads closer to being practical without a cloud bill.
What to watch: whether fintechs like Ramp turn model routing into a wedge for selling their core products, and how far on-device speculative decoding pushes small models into real agentic workloads.
Should every fintech with an AI story become an inference middleman, or is the router market already crowded? Tell us in the comments.
