Show HN: Optimize and serve models with Fable quality at half the cost

7/27/2026 at 12:08:55 AM

Local models need to be tuned to work well so this looks useful. Seems to be for general purpose model serving. I’ve been using https://github.com/adrianco/retort to run experiments for coding models across 13 different programming languages to see which frontier and local models work.

by adrianco

7/27/2026 at 12:14:46 AM

That's cool, thanks for sharing!

by SilenN

7/27/2026 at 12:05:41 AM

Not sure I get it. The model you're improving is local? If so how do you even calculate cost compared to an API

by jack_pp

7/27/2026 at 12:21:00 AM

Open source models.

wmo routes requests between frontier models and open source models that continuously train using Tinker. As the smaller models improve, more traffic gets routed to them.

Calculating cost is just tokens in/out.

by SilenN

7/30/2026 at 8:16:04 PM

What are the costs to train and use the Tinker models?

by handfuloflight

7/30/2026 at 9:01:26 PM

Expensive, in the thousands. We have our own infra in house and are working on bringing these costs down

by SilenN

7/30/2026 at 9:14:11 PM

But in the thousands can pencil out if you're saying tens of thousands over calling the frontier models, no?

by handfuloflight

7/29/2026 at 1:38:39 AM

Interesting approach. What does the cold-start phase look like for a new agent? How many traces or runs do you typically need before the router has enough signal to safely offload tasks from the frontier model??

by anshad2u

7/30/2026 at 8:33:34 PM

Technically 0 because a) it ingests your already existing traces and does an initial training run b) in the app we'll have pre-trained routers you can start with that will then learn over time

by SilenN

7/27/2026 at 12:26:51 AM

The title is misleading. This is model routing, not distillation.

by surround

7/27/2026 at 12:31:53 AM

Fixed formatting which will help with readability. We do routing, distillation, and token compaction.

by SilenN

7/27/2026 at 12:07:46 AM

Cool project

by digitaltrees

7/27/2026 at 12:31:17 AM

Thanks :)

by SilenN

7/27/2026 at 12:07:41 AM

Cool idea! How do you guarantee privacy?

by yiyingzhang

7/27/2026 at 12:22:12 AM

It's open source!

We do have a platform we'll be launching as well to manage training + serving for you which will require more diligent privacy guarantees.

by SilenN

7/27/2026 at 12:03:43 AM

Excited to play with this.

by rglover

7/27/2026 at 12:31:28 AM

Let me know if you have any questions!

by SilenN

7/27/2026 at 12:15:43 AM

The absolute best way to prove this works is by releasing a model that was fine-tuned with this method and then showing benchmarks depicting the improvement delta between the base model and the fine tuned one.

The work is not done. Then release it to the masses and wait a few days for the actual real world anecdotes.

Until then, this is noise.

by Art9681

7/27/2026 at 12:22:29 AM

Valid criticism. Happy to answer any qs. We're still working on solidfying results.

by SilenN

7/27/2026 at 12:24:08 AM

Ok, I think it is in your interest to wait until you have more to show, and we'll be happy to help you with reposting it once it's ready.

Waitlists are against the Show HN rules (https://news.ycombinator.com/showhn.html), and you're likely to get a lot of community pushback if you post before there's enough substance for users to sink their teeth into.

Edit: we eventually got a more substantive writeup from OP so I moved that text to the top and re-upped this thread.

by dang

7/27/2026 at 12:30:25 AM

Thanks for the heads up, removed mention!

by SilenN

7/27/2026 at 12:20:15 AM

Benchmarks are the ultimate consolidation of halnons razor.

by irishcoffee

7/27/2026 at 12:21:42 AM

[flagged]

by teravor

7/27/2026 at 12:22:23 AM

"Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something."

https://news.ycombinator.com/newsguidelines.html

by dang

7/27/2026 at 12:19:36 AM

Yeah, this is just slop. No benchmarks, no concrete case studies, just some vibecoded "platform" to finetune models on your own traces.

Which is an idea that has some value, but also some weaknesses. And this implementation of it isn't forthcoming with that concept. You have to really dig in to understand what they're even talking about.

by Reubend

7/27/2026 at 12:25:28 AM

Happy to answer any qs.

by SilenN

7/27/2026 at 8:10:47 AM

[flagged]

by renezander030