Video wird geladen...
Video konnte nicht geladen werden
Introducing Liquid Inference by Architect, a new LLM model router where inference providers compete dynamically to offer lowest-cost tokens. Providers onboard in minutes, not weeks. Buyers have full price visibility and receive 20% of referred fees in free inference. More info: Liquid Inference can be used in place of... show more
38,511 Aufrufe • vor 18 Tagen •via X (Twitter)
37 Kommentare
We’re excited to launch Liquid Inference with some of the fastest-growing token producers in the ecosystem: • Engy @engyai • Pearl Research Labs @prlnet • AkashML @akashnetAI • Hyperbolic @hyperbolic_labs • Boundless @BoundlessHQ • Zro @zroai_
Sign up to get started now:
The hard part of an exchange is rarely price discovery, it is that the good has to be verifiable at delivery. Inference is the unusual case where it is not. Nothing in the output tells a buyer which weights produced it. So competition runs on the one attribute that is observable, and the attribute that sets value is not. The failure mode is not fraud, it is quantization. Serving an int4 build of a listed model is cheaper, looks identical on easy prompts, and degrades precisely on the hard ones, which are the prompts anyone is paying real money for. That makes the interesting problem here surveillance rather than matching, and the analogue is on your side of the fence already. Exchanges do not ask members to self report execution quality, they sample. Blind probes against listed providers, published per provider, are what would make a price on this good mean anything. Separate note: a rebate that lands with whoever controls routing is the structure best execution rules were written for.
Excited to be one of the launch partners! We’re building toward a future where compute is easier to access, manage, and trade. Congrats on the launch. Looking forward to what we’ll build together!
Compute markets 👀
@vikramcantsingh so fucking cool, was just talking to @lzminsky about this frontier last night. big congrats also speculated about this a few weeks ago on twitter. what do u think about the post it’s replying to?
I think the frontier model companies are generally trying to capture as much of the customer UX as possible, and will continue delegating the router resales to other companies that can optimize dynamic pricing. I could see that changing if their hands are forced by enterprises migrating to routing solutions.
@Architect_Fi Liquid Inference is a very smart name! Will give it a try with some harness tools today.
Liquid Inference tests the network not just tokens. You save money but you add latency variance to your pipeline
Spend 15 minutes on our new site, and you'll soon better understand the AI economy. Please give it a try. @BrettHarrison @Architect_Fi
Huge! Congrats Brett and team!
Cool man. Congrats! Hope to see you crush it.
Providers competing on price means the lowest cost wins every single time
Exciting launch!
1. How is it different from OpenRouter? 2. How does KV work? 3. How dynamic is the pricing?
how do buyers know a cheaper provider isnt serving a quantized version of the same model?
Woah this is awesome!
price transparency in model routing solves the wrong problem. buyers optimize for cost per token but inference quality varies wildly across providers on the same model. routing needs accuracy metrics per provider or the lowest price becomes a quality lottery.
Real-time price competition is the interesting piece; I’d watch how consistently it beats static routing once provider costs shift.
Sick!
when everyone's competing on tokens/m the only thing left to lose is quality. curious how you catch a provider degrading throughput or p99 latency once the price war starts - that's where the actual cost of 'dynamic' lives
sick @grok how good is the privacy / data retention / data usage
price discovery for tokens is interesting, but the cheapest quote for the "same" model is often a heavier quant or a shorter context. do quotes carry quant/precision + measured tps/ttft? otherwise the auction quietly optimizes for worst-acceptable quality.
How do you make sure providers are serving exactly what they say they are serving? You could have malicious it heavily quantized versions or simply a very slow or low uptime provider
Very sick!!!
bringing an orderbook mindset to token inference is so clean! definitely plugging this into cursor
This is sick af Brett!
Hashrate got a spot market. Tokens were next. Congrats Brett & Team! Excited to try it out.
Love this dynamic price competition for inference with real exchange mechanics is such a clever take and 20% back in free inference is a great incentive
This is very cool. We will be testing!
an exchange for inference prices is delightfully literal
A 40-step agent run that hops to the cheapest bidder every step also hops away from its prompt cache every step. The cheapest token is often the one you already paid to cache. Does the router pin a session to one provider, or is that on the buyer? #AIAgents #LLM
Clever approach to making inference pricing more open
@andrewdsouza architect routes inference buyers to competing GPU providers
@grok routing on lowest price only works if quality is flat across providers. It isn't: the cheap one quantized harder, so your eval drifts under you. Bet the 10th-percentile provider fails 3x the median.
Price-first routing is a commodity bet: fine while the models are interchangeable, awkward the moment a buyer cares which one answered.
Lowest-cost routing gets tricky when two providers serve the "same" model at different quantizations. How do you keep output quality consistent across providers?
