Building research-led capabilities for product and enterprise teams. Synthetic data · Custom training · Evals Solving problems generic APIs cannot solve.
Joined October 2026
- Tweets20
- Following8
- Followers189
- Likes17
Pinned Tweet
Today we introduce Mercity.
A research-first AI firm building at the intersection of synthetic data, custom training, evals and safety
We believe that the future of safe enterprise AI lies in enabling research led capabilities
Here's our work so far :
mercity.ai/
Today we're open-sourcing Heckler.
Heckler uses simulated users to test your agent.
It is purpose built to push your agent's system rules to the limit and tells you where exactly your agent messed up.
It not only catches outright jailbreaks, but also subtle policy violations and rule breaking behaviour.
This makes testing your agents and prompts before they hit production way easier and simpler!
Try it here :
heckler.mercity.ai/
Today we introduce Heckler by @MercityAI
Heckler stress tests your AI agents with simulated users and shows you exactly where your agent breaks the rules you gave it.
We ran 3 agents (DeepSeek V4.1 Flash, GLM 5.3 and GPT-6 Luna) on Heckler and over 80% of chats broke a policy.
Our Latest Research:
We pretrained four 1B-parameter models on 6B FineWeb tokens to test two ideas from frontier MoE labs, Kimi Delta Attention and LongCat n-gram embeddings, at small scale.
We then compared grafting, pruning, 4-bit quantization, and downstream NER fine-tuning
We then operated on trained checkpoints
Removing layer 0 cut the score by 1/3rd while most layers past layer 8 could go almost for free
4-bit quant cost under a point
Attention could not be moved between models without collapsing
Full research log: mercity.ai/research/pretrain…
We REMOVED upto 10 WHOLE layers from Qwen3-8B-Base using LaCo (Layer Collapse).
Upon deleting only 6 layers, the model got 16.7% smaller, keeping 78-94% of its reasoning scores at a 1.68x perplexity ratio.
Code and pruned weights are up on our Hugging Face.
Here are insights from our research :
The paper's Llama settings didn't transfer to Qwen3
A 0.65 similarity threshold sent MMLU to random, 0.85 worked
Merging near the output broke the model, so we stopped at layer 28
Merging LayerNorms gave gibberish, so we left them alone
There is probably an element of how many tokens the model is trained on! We will explore this in the coming times!
Find the full research log here:
mercity.ai/research/laco-lay…
We build custom data, models and evaluation for problems an API cannot solve.
If that is the problem in front of you, our DMs are open, or start a conversation at mercity.ai
Find out more about our Models, Research, Datasets and work here :
mercity.ai
huggingface.co/Mercity
github.com/Mercity-AI