Rendered at 09:20:48 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
d2p 11 minutes ago [-]
> and use your traffic to (opt in) train you a model.
Is there more info on this? I'm curious exactly what it is. Is it fine-tuning/LoRA on some base model? Don't cloud providers encrypt reasoning now - does that prevent this?
nejch 34 minutes ago [-]
A part of me wishes the open source community would focus making research and industry-backed initiatives like the vLLM Semantic Router rock solid. Then I'd spend less time every month checking if this or that new model router has differentiating over vllm-sr :)
At least for open source inference, it seems like there's healthy competition centered around vllm/sglang, but 2026 seems to be for model routers what 2025 was for agent harnesses.
Areibman 11 hours ago [-]
Could you say more about how caching works? One major advantage of sticking with a single model is saving money on cached input tokens. I'd imagine if you swap between a bunch of models, you may improve performance but cost would would balloon out of control
SilenN 11 hours ago [-]
The trick is to rarely switch, or switch at task boundaries. Often the conclusion of routing is actually "this one model is actually at the pareto front for this task, just use it always".
cameronh90 9 hours ago [-]
But then it's better to just not have a gateway switch models at all.
Just have the harness able to choose which model its sub-agents use, then tell it how to split up tasks and which models to use when doing so.
SilenN 9 hours ago [-]
That is another way to do. Or we can automatically figure out which models the subagents should be using for you. And update them as new models come out and the work your subagents do changes. More than one way to skin a cat.
Open source and no markup is the right default for a gateway. The caching question above is the one I would want answered before swapping models though.
SilenN 8 hours ago [-]
Ans: we rarely switch, often times it's just a "switch to using this model for your agent"
ceroxylon 8 hours ago [-]
>The gateway adds under 1 ms for BYOK requests
Amazing! Really brilliant idea, thank you for sharing this project. There is so much ground to cover in the LLM gateway / routing / reporting world, and this is a great start. The Tinker implementation is my favorite part, fine tuning is much better than a sea of context files.
kfallah15 8 hours ago [-]
Thanks! We are going to add continual RL via Tinker soon too
cheema33 9 hours ago [-]
I have not tried it yet. Is it similar to LiteLLM? If so, what sets it apart?
kfallah15 9 hours ago [-]
Router and model optimization from traffic is the main differentiator
SilenN 9 hours ago [-]
Also a hosted marketplace, not just BYOK
forgetme2020 6 hours ago [-]
what's the business model here. How does experiential labs make money
Look at the Intelligence features in the Enterprise plan:
* Per-prompt model optimization
* Caching
* A model you own, trained on your traffic
kfallah15 5 hours ago [-]
yep, it will be through enterprise licenses and our own hosted platform built on the repo
rdslw 2 hours ago [-]
the business model, I suspect, is classic rug-pull in some time after building user base.
proof: boldly claiming being open source in literally first sentence, while cowardly hiding on-by-default telemetry (WTF??) in truly last paragraph of readme.
sorry to sound harsh, but this is typical old era playbook here.
in the era of AI, fortunately, such products has much lower value. people and VCs didn’t yet tune to it.
SilenN 2 hours ago [-]
Telemetry is off by default. PostHog is for usage analytics on the open source repo. Audit it if you're skeptical.
We make money off enterprise licenses and hosting models.
I have strong reason to suspect you either can't read or are a bad actor.
gpiechnik2 1 hours ago [-]
great design! i love it
SilenN 41 minutes ago [-]
Thank you!
swthbht 6 hours ago [-]
Very cool. Does your gateway decide effort levels as well? Or just models?
SilenN 6 hours ago [-]
Yep! One interesting example is often Opus 5 on low reasoning ~= Opus 5 on high reasoning.
sangwook 6 hours ago [-]
What online signal recalibrates simulated rankings against actual task success? Also do you have a plan to support semantic caching at the router level?
kfallah15 5 hours ago [-]
For the online signal, we use a LLM judge with a rubric calibrated offline by the user via TUI. UX of the calibration is a major focus area. Semantic caching is interesting, open to supporting it but not currently planned.
0xbadcafebee 8 hours ago [-]
You started it a week ago? I look forward to checking back in 3 weeks when you've exited for $1B
Thanks for the positivity tyre! If you look at our git history, we pivoted and only started building the gateway recently. Before that we were building research infrastructure that now powers the intelligence features we provide.
rdslw 2 hours ago [-]
impressive only if using pre-gpt era assumptions about saas/products/software.
unfortunately a small team can reproduce it in two months, which greatly lowers value of it.
we, as a collective, have to change our value-judging logic and tune it to post AI world.
SilenN 7 hours ago [-]
See you soon
23david 11 hours ago [-]
Super interesting and congrats on the release. Curious if you initially had this in Python and then rewrote in Rust?
SilenN 10 hours ago [-]
Yep! If you look at the commit history that's exactly what happened.
Is there more info on this? I'm curious exactly what it is. Is it fine-tuning/LoRA on some base model? Don't cloud providers encrypt reasoning now - does that prevent this?
At least for open source inference, it seems like there's healthy competition centered around vllm/sglang, but 2026 seems to be for model routers what 2025 was for agent harnesses.
Just have the harness able to choose which model its sub-agents use, then tell it how to split up tasks and which models to use when doing so.
Amazing! Really brilliant idea, thank you for sharing this project. There is so much ground to cover in the LLM gateway / routing / reporting world, and this is a great start. The Tinker implementation is my favorite part, fine tuning is much better than a sea of context files.
Look at the Intelligence features in the Enterprise plan:
* Per-prompt model optimization
* Caching
* A model you own, trained on your traffic
proof: boldly claiming being open source in literally first sentence, while cowardly hiding on-by-default telemetry (WTF??) in truly last paragraph of readme.
sorry to sound harsh, but this is typical old era playbook here.
in the era of AI, fortunately, such products has much lower value. people and VCs didn’t yet tune to it.
We make money off enterprise licenses and hosting models.
I have strong reason to suspect you either can't read or are a bad actor.
So, two months. Still impressive!
unfortunately a small team can reproduce it in two months, which greatly lowers value of it.
we, as a collective, have to change our value-judging logic and tune it to post AI world.