The compute efficient layer for AI inference
Austin, TX
Based in United States
- Tweets 240
- Following 30
- Followers 178
- Likes 194
Our ZLM for IAB classification beat GPT-5.4 nano in 66% of 10,000 blind tests, about 10x faster, at about a tenth of the price.
Built for publishers, SSPs and web-data companies.
zerogpu.ai/benchmarks/iab-cl…
ALT ZeroGPU card for zlm-v1-iab-classify-edge. Reads a web page and says what it is about, so the right ads can run next to it. For publishers, ad platforms and web data teams. Benchmark figures: 66% win rate against gpt-5.4 nano, 48 ms p50 latency, against ~1,900 ms, 0 hallucinated labels, against 1,796.
2
1
3
51
Check out our model catalog! 😇
We run open weight models and our own task-specific small language models more sustainably and more cost-effectively.
Our hybrid inference cloud pairs AI tasks with idle compute in laptops and phones
40
could be using a SLM!
We build small and nano models for routine AI tasks like these that use way less energy than frontier models like ChatGPT or Claude.
Cuts down on AI costs and is more sustainable bc it doesn't need an expensive data center or GPU.
1
167
100% - take advantage of the right model to reduce your inference bill.
Most of the stuff AI is used for, classification, summarization, data extraction, could be run by a way smaller model for way less.
That's why we're running those SLMs & open-weight models on everyday devices to reduce costs.
5
We'll be highlighting what you can do with our ZLMs: ZeroGPU Language Models.
Small models trained for one job each, like classifying, moderating and extracting from text. One job a day, with the industries it's built for.
docs.zerogpu.ai/docs/model-c…
ALT ZeroGPU card for ZLMs. Small AI models that each do one job well, instead of one big model doing everything. For any team running AI at scale. Benchmark figures: 5 zlms in the catalog today, 50+ languages, no translation step, $0.02 per 1m input tokens.
1
30
Limelight enriched 2.88M prospect rows on ZeroGPU in six weeks.
9.2x cheaper than Claude Haiku, 3.2x cheaper than Gemini 3.5 Flash-Lite, $12,739 saved. Signup to production in 34 hours.
Case study: zerogpu.ai/case-study/limeli…
ALT Limelight case study card: 2.88 million prospect rows enriched on ZeroGPU in six weeks, 9.2x lower cost than Claude Haiku and 3.2x lower than Gemini 3.5 Flash-Lite, $12,739 saved.
1
43
You can now access our ZeroClick agent storefront on Locus Pro.
Agents can find our specialized small language models and pay from a Locus wallet to tap into cost-effective AI inference.
No signup, no API key. Just a better use of your AI spend.
All ZeroClick-powered Agent Storefronts are now listed as verified services on Locus Pro 🔥
If you want to sell your product to AI agents, set up an Agent Storefront with ZeroClick!
2
3
105
Not every task needs a frontier model.
More than 40% of OpenRouter spend is already SLM-ready. By switching to the right small language model, you can reduce latency, cut costs by 5x.
We have more models coming soon. Stay tuned!
zerogpu.ai
2
3
74,756
Frontier models are great for high-level reasoning, but you don’t always need those expensive models.
We pair the most common AI tasks with idle compute.
Get started today.
zerogpu.ai
2
1
3
31
ZeroGPU is now an @nvidia Inception member!
This gives us deeper access to the NVIDIA platform, tooling and ecosystem as we work to cut the cost of AI inference.
Run high volume workloads on our small and open-weight models on-edge.
- 5x savings
- 10x reductions
- the same accuracy at a fraction of compute
If you’re ready to start cutting your AI spend, check out our SLM catalog.
3
1
3
93
ZeroGPU is now on LangChain 🦜
If you're running a @LangChainAI agent, a lot of what that agent does on every turn doesn't need frontier reasoning.
Point those tasks at specialized small language models instead and cut your inference bill.
1
2
4
99
Drop into any pipeline already written against OpenAI's moderation models to instantly save costs, while maintaining accuracy across responses.
Full benchmark and docs: zerogpu.ai/benchmarks/modera…
1
2
20
Moderation sits in front of every response your app serves. At frontier prices that volume gets expensive fast.
So we trained an SLM to manage those tasks while running on edge.
Save frontier model work for where it counts - use ZeroGPU for those tasks you run constantly.
1
2
68
Moderation runs in front of everything your app serves.
Every message, every generation, every time. That's a lot of volume to be paying frontier prices for. So we built a new SLM for moderation, designed to run at the edge, using less compute to get more done.
We benchmarked our new zlm-v1-moderation-edge model head-to-head against OpenAI's omni-moderation-latest. Take a look at the results.
The takeaway? Save your frontier models for tasks that require deep reasoning. Use ZeroGPU for the repeatable work that runs a million times a day.
Full benchmark + docs linked in the comments ⬇️
2
3
8
465
We’re excited to partner with the ZeroClick team on this!
Let agents buy AI inference directly.
Get started: zerogpu.ai
5
44
Most production AI work can be done for less.
Now through @zeroclick, agents can purchase inference, tapping into our specialized nano models & SLMs. Let your agents save frontier model tokens for your highest-level reasoning work.
Use ZeroGPU for everything else.
We're excited to partner with @ZeroGPU_AI, enabling AI agents to directly buy specialized inference 🤝
zeroclick.ai/blog/zerogpu-sp…
1
2
10
181
Most teams are paying more for frontier models for simple tasks that don't need that level of complex reasoning.
That’s why we’ve built a compute-efficient layer for AI inference.
Tap into open-weight models through an OpenAI-compatible API or via our Claude Code Plug-in, with support for GLM-5.2, gpt-oss-120b, Qwen3, Llama 3.1, DeepSeek & Moonshot Kimi K2.
Get started: zerogpu.ai
11
10
258
ZeroGPU Router plugin is now on OpenClaw 🎊
Your host model is doing a lot of work you shouldn't be paying frontier prices for. Now with our router, your host model only takes on your highest-level reasoning tasks. The repeatable work runs on our SLMs and nano models.
20+ skills available. Three commands to install
2
1
2
62
Huge credit to the @Kimi_Moonshot team for releasing the weights.
Though we are focused on our edge network, open-source is a good thing!
Releasing the model weights and technical report of Kimi K3.
Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.
New model architecture: 2.5x the intelligence per unit of compute, not just more params.
Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.
Model weights: huggingface.co/moonshotai/Ki…
Tech report: github.com/MoonshotAI/Kimi-K…
Tech blog: kimi.com/blog/kimi-k3
1
3
85
