Editorial illustration for Aikido Security Releases Altar-1, A 328GB Model Pruned From GLM-5.3
Aikido Security Releases Altar-1 Security Model
Aikido Security put its first open-weight security model on Hugging Face this week. Altar-1 is a pruned-down version of Z.AI's GLM-5.3, shrunk from 753 billion parameters to 328 GB, small enough to run on a single node of four NVIDIA H200 GPUs with vLLM. It's the engine behind Aikido Machine, the company's autonomous pentesting appliance built for on-prem and air-gapped networks, the kind of environment where sending code and vulnerability data to an outside API isn't an option.
That's the pitch: banks under data-residency rules, OT operators with no internet route, anyone who can't let source code and unremediated findings leave the building. Open-weight models get around that by staying local. But running a model like GLM-5.3 in-house creates its own problem.
It's a mixture-of-experts model, and MoE architectures have to keep every expert loaded even when only a handful get used per token. Add in the long-running context that security agents build up during a scan, and GPU memory fills fast with weights and KV cache fighting for the same space.
Aikido Security has released Altar-1, its first open-weight security model. It is a compressed version of Z.AI’s GLM-5.3, built to run inside infrastructure the customer controls.
Why this matters
Altar-1 is a bet that security teams will trade raw model quality for the ability to keep pentest data off someone else's servers. Running on 4x H200s isn't cheap, but it's a fixed, known cost compared to sending network topology and vulnerability data through a third-party API. For founders building in regulated or air-gapped sectors, that tradeoff might matter more than benchmark scores.
We'd want to see how Altar-1's pruned, quantized version actually performs against the full GLM-5.3 on real pentesting tasks before calling this a viable alternative to closed frontier models, not just a compliance workaround. Publishing weights on Hugging Face is a genuine signal of confidence, and it lets researchers verify Aikido's compression claims instead of taking them on faith. Watch whether other security vendors follow with their own open-weight, on-prem models, and whether Aikido reveals what got cut from GLM-5.3's expert layers to hit that 328GB target.
That detail will tell us how much capability was actually sacrificed for control.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives AI Chip Designers on When AI Starts Designing Its Own Hardware
From Archives AI's High Cost: USD 856B in Compute, IPO Plans Strain Sectors
Common Questions Answered
How much was Altar-1 pruned from the original GLM-5.3 model?
Altar-1 was significantly reduced from Z.AI's GLM-5.3, which originally had 753 billion parameters, down to 328 GB. This pruning makes it small enough to run on a single node of four NVIDIA H200 GPUs with vLLM, enabling deployment in resource-constrained environments.
What is the primary use case for Altar-1 in Aikido Machine?
Altar-1 serves as the engine behind Aikido Machine, Aikido Security's autonomous pentesting appliance designed specifically for on-premises and air-gapped networks. This design allows security teams to conduct penetration testing without sending code and vulnerability data to external APIs, which is critical for regulated environments.
Why would security teams choose Altar-1 despite potential quality tradeoffs?
Security teams may prioritize Altar-1's ability to keep sensitive pentest data and network topology off third-party servers over raw model performance benchmarks. Running on 4x H200s represents a fixed, known infrastructure cost compared to the ongoing risks of sending vulnerability data through external APIs, making it particularly valuable for regulated or air-gapped sectors.
What hardware requirements are needed to run Altar-1?
Altar-1 can run on a single node equipped with four NVIDIA H200 GPUs using vLLM. This relatively modest hardware requirement for a 328GB model makes it feasible for organizations to deploy it within their own infrastructure without requiring extensive computational resources.
Further Reading
- Aikido Security Releases Altar-1: An Open-Weight Security Model Pruned From GLM-5.3 to 328 GB - MarkTechPost
- Aikido ships Altar-1, a 328 GB GLM-5.3 prune for on-prem security workflows - RuntimeWire
- Aikido Security Unveils Altar-1 Open-Weight AI for Cybersecurity Defense - Cyber Security News
- Aikido Altar Brings GLM-5.3 Security Workloads On Prem - Security Boulevard
- Aikido、オープンウェイトのセキュリティ向けAIモデル「Altar-1」を公開 - gihyo.jp
