model routing protocol and runtime, and a package for Pi by @trydotworks https://xitter.cf/t.co/cM4OR5Wqb8
Sweden
Joined June 2026
- Tweets41
- Following22
- Followers292
- Likes10
Pinned Tweet
Support for using any model in Codex via role-model is in beta and launching next week.
If you can't wait, its already live in Pi with the pi-role-model extension.
lmao Codex now proudly tells me I can use DeepSeek, after i configured model routing with @rolemodeldev
role-model retweeted
A model router that learns how to route around your codebase based on your requests.
Runs silent evals in the background. Extracts learnings.
Locally, your machine.
I've been working on this for the past 2 months.
Coming soon in @rolemodeldev
role-model retweeted
Flash models from DeepSeek, GLM and Qwen are creating a problem: they overperform on benchmarks which makes it hard to know when and how to use them.
Even in this benchmark in @rolemodeldev which uses a judge to grade the solutions from each model side by side and set scores relative to each other, V4 Flash outperforms V4 Pro and Terra High.
It certainly isn't better for long-running work because it eventually loses the plot and stops following instructions.
It reinforces the idea that benchmark scores cannot be used for model routing; instead we must run replay and evaluations on actual traces from work you have been doing in your repo.
That is coming in a soon to be released version of role-model.
role-model v.0.0.11 is now out and supports effort-aware routing which means you can route between Low, Medium, High and Max variants of the same model for more granular speed, cost and quality trade-offs.
Go get it: github.com/try-works/role-mo…
role-model retweeted
This AI review rated @rolemodeldev 8.5/10 so you know it's good aiindigo.com/blog/role-model…
role-model retweeted
Stop thinking. Let @rolemodeldev decide.
Can't decide whether to use DeepSeek V4 Flash or Pro? Let the router decide for you.
Zero effort after setup.
Here's how role-model helps you make DeepSeek models into capable designers:
DeepSeek Harness now supports @paper via the plugin dsh-paper-design I put together.
This plugin has 1:1 parity with the Cursor plugin for Paper. As you may already know, working with Cursor and Paper is like magic for design.
In a first experiment, DeepSeek V4 is able to (somewhat) successfully rebuild its own web chat interface despite not having access to vision (image input) capabilities. I guess design is just code after all?
What is mainly lacking is the half-done DeepSeek logo that looks more like a lobster than a whale. It also chose to feature somewhat outdated models, R1 and V3.
The lack of vision is easy to overcome - just run @deepseek_ai Harness with the role-model router @rolemodeldev and add a vision-capable model like GPT or Kimi, and this issue is fixed. Thanks to Paper, dsh-paper-design and role-model router, DeepSeek models with DeepSeek Harness are capable of doing great design - better than this first attempt for sure.
role-model retweeted
the flat line at around 98% is the cache hit rate in @rolemodeldev router
it's actually higher than that, usually >99% per request. the reason the graph shows 98% is that it averages the cache hit rate per request, and hit rate is always lower for the first few requests in a new session since every part of the context is new, or when you start doing something new in the same session.
so yes, the role-model router helps you maintain a cache hit rate above 99%.
Heads up, role-model works perfectly with DeepSeek Harness. Cache hits reaching >99% through the router.
A DSH plugin for role-model is coming soon, and will offer richer routing metadata and task taxonomy, for detailed routing decisions.
role-model v0.0.10 is out with major new features:
- Updated UI
- Message graph
- Knowledge store and knowledge workers
- Support for exporting traces for RL in Prime Intellect
And much more is coming in the next week.
github.com/try-works/role-mo…
Updated data insights dashboard on where your requests are routed and what your models are actually doing - coming next week in role-model.
Learn more about role-model on our new site: role-model.dev/
