Models and infra for AI code generation.
- Tweets 44
- Following 9
- Followers 763
- Likes 38
Relace retweeted
Relace served 7.2T tokens out of the 146T total that passed through OpenRouter last week.
We've achieved this by using a pricing algorithm that maximizes our token share across 9 different models, optimizing each of our model deployments for throughput.
On the open market inference has become another form of high frequency trading.
5
3
21
5,492
If you want to spend all day working on inference optimization, we're hiring. relace.ai/about-us
Relace is now also the cheapest provider of DeepSeek v4.1 Flash on OpenRouter.
Agent traces are 96%+ input tokens, so GPU utilization is dominated by prefill time. However, most of those tokens are not relevant for the next response.
The encoder/decoder split allows you to use a lossier, cheap submodel for prefill without affecting the performance of decode (i.e. quality of responses).
It also means you need to shuttle KV between GPUs a lot, which is why they redesigned attention (again) and made the KV 4x smaller.
Fully taking advantage of the theoretical gains is a deep engineering problem that DeepSeek left as an exercise for the reader.
2
1
3
1,090
Today we did 37% of Deepseek v4 Flash, 25% of GLM 5.3 Flash, and 1.6% of Kimi K3 traffic.
1
1
3
1,163
A partnership of the white on black logos :)
Relace is live on @Vercel AI Gateway as the cheapest Deepseek v4 Flash Provider!
Inference is highly inefficient, and there still gains to be squeezed out of every layer in the stack. We're bringing those gains directly to the consumer.
2
3
702
We've spent a lot of time optimizing inference for speed -- 10k tok/s Fast Apply and a 50k tok/s Compact model.
We're now focusing on making our favorite set of open weight coding models as economical to serve as possible.
After only 2 weeks, we're at the top of OpenRouter for Deepseek. More to come soon!
Relace is now sustainably serving Deepseek v4 Flash almost 2x cheaper than anyone else on OpenRouter.
We're working on optimizations to bring this down even further.
If you're offering a free plan for your coding agents with Flash, reach out to us. We'd love to support you too!
2
1
7
1,118
We just dropped the first token gift cards
Open-source models have gotten really good, but running them is still expensive. That's why we’re launching our inference service to make them cheaper—and releasing 10 Kimi K3 packs to the community: $500 in credit for $50 each. Here’s to making open models more accessible! packs.relace.ai/?utm_source=…
2
606
Relace retweeted
What I love about AI: remixing ideas becomes so easy
Inspired by @marina_uiux's work. built using @relace_ai's blazing fast Jacq and @QuiverAI 💐 🕊️
The text is from @pmarca's It's Time to Build
The page plays music and spawns flowers around the @a16z sf office
11
16
286
12,186
Try it out for yourself on any model (Fable, GPT 5.6, Grok 4.5) at jacquard.dev or check out docs.relace.ai to integrate Relace Compact into your own product!
We found compacting agent context on cache misses reduces token costs by over 50%.
The graphic is a real Fable 5 trace from our CTO, where we simulate cost would before/after compacting after periods of over 5 mins of inactivity.
By keeping context within 128k-256k tokens, spend goes from being quadratic to linear in the number of turns.
The only problem is self-summary takes 1-2 minutes on big traces! We trained an absurdly fast compact model to get the cost savings without harming the user experience.
Relace Compact is a modified LLM designed for token-level classification, which means we can run it at 50k tok/s and an order of magnitude lower than cache read cost ($0.20/million input and $0.20/million output tokens).
Try it out in Jacq, or visit our docs to integrate into your own product!
1
6
2,011
First class plugins in Jacq are becoming a window into all your daily tools accessible from one place.
2
283
One of the many first class integrations, we're bringing into Jacq. Try it --> jacquard.dev
1
338
Relace retweeted
You lose a lot of flexibility when a cloud agent is quarantined in its VM until you sync manually.
Jacq packages a toolbox binary that turns each device (phone, laptop, desktop) into an API, and the agent dynamically decides what connected device to execute tools on.
This means if my laptop goes offline, Jacq can autonomously decide to keep working on other things in the cloud until I'm back.
Long term it's always better to give the model more choice.
1
4
13
1,279
Relace retweeted
A lot of coding products treat threads as disposable, and you end up having to get your agent up to speed with the same info over and over.
We trained a compaction model for Jacq that runs at 90k tok/s to make resuming with context from other threads instant.
1
3
8
1,038
For 18 months, Relace has been training small, specialized models that make autonomous coding products smooth.
Jacq is built on everything we've learned and showcases some new models we've worked on recently.
It uses a compaction model under the hood that runs at 90k tok/s, fast enough to condense a 250k+ token thread in 2-3s. This means you can stay on the same thread indefinitely, or instantly fork from the same starting point.
Autopilot mode is actually a small classifier model that identifies unsafe tool calls. The agent stops bothering you about routine stuff, and is forced to work around or ask you about things you might regret.
Small models are best when they feel invisible and just work. Let us know what you think, and DM us if you want API access to these models!
Today we're launching Jacq. A coding agent built together with the small models we've been training at @relace_ai for the past 1.5 years.
It runs entirely in the cloud, and decides when to pull context from any of your connected devices.
It uses all the software you already live in: Slack, Linear, GitHub, email, etc.
Plus, threads are now durable — a real record of how work happened. Just drag them into a new chat to get context for your next task.
3
1
16
4,099
🚀
Building a website used to take months. I created 5 in 10 seconds. Meet Fimo.ai, an AI-native CMS where your website lives, collaborates, and evolves with AI.
Existing vibe coding tools are great for building apps, but websites have specific requirements. We’ve spent the last 10 years building Strapi, the most popular open-source Headless CMS. We know what is important for users: translate content, reuse assets, optimize SEO/GEO, update content, A/B test, personalize pages, schedule releases, review workflows, etc.
4
1,250