Aleph Alpha's open-weight English-German mixture-of-experts model for self-hosting, released under Apache 2.0 in October 2026
Altar-1
New· 4Defensive security model from Aikido Security, a pruned open-weight version of GLM-5.3 for on-premises testing
About Altar-1
Aikido Security, a security company, announced Altar-1 on 21 September 2026 as its first open-weight model. It is a compressed version of Z.AI's GLM-5.3, meant for defensive security tasks inside systems the customer controls itself, air-gapped networks included, where a hosted model is not an option.
Where it is used
Altar-1 runs inside Aikido Machine, an autonomous penetration-testing appliance that the company installs on premises. Aikido says it is extending the approach to its Deep PR Review, Code Security Audit and Aikido Attack products.
How the model was shrunk
Aikido took a community AWQ INT4 quantization of GLM-5.3 as its starting point and pruned experts with Cerebras's REAP method. Calibration drew on traces from Aikido's own pentesting harness along with coding, tool-calling, reasoning and multilingual text, and the company states that no customer data went into it. In every expert layer, 168 of the 256 routed experts remain, and eight are still chosen for each token. The model card puts the total at about 504 billion parameters, with roughly 40 billion active and routed experts stored at 4 bits. The full-precision parent needs 1.51 TB of storage, against 328 GB for Altar-1. Aikido says one node with four NVIDIA H200 GPUs can serve the model through vLLM, and its example command sets a context of 131,072 tokens.
What Aikido measured
Aikido's own test used 32 known vulnerabilities spread over 30 repositories. It reports average recall of 60.4 percent per run, compared with 61.5 percent for the quantized parent and 65.6 percent for full-precision GLM-5.3, and says Altar found 23 of the 32 at least once. The company cautions that the test measures targeted rediscovery rather than blind discovery or exploit validation.
License terms
The weights can be downloaded free under the GLM-5.3 license, which permits commercial use and modification. One condition applies to large providers: a company with a Model-as-a-Service business and revenue above a stated threshold in any 12-month period must first clear a Z.AI security review. Reuters reported the launch on 21 September.
TrustList summary
Altar-1 is the first open-weight model from Aikido Security, a pruned and 4-bit compressed form of Z.AI's GLM-5.3. It is meant for defensive security work on infrastructure that customers control, including air-gapped networks.
Main features
- Aikido Machine: the model powers Aikido's on-premises autonomous penetration-testing appliance.
- Wider use: Aikido says the same method is being carried over to Aikido Attack, Code Security Audit and Deep PR Review.
- Self-hosted serving: runs on one node of four NVIDIA H200 GPUs with vLLM, with an example set to a 131,072-token context.
- Open weights: free to download, and the GLM-5.3 licence permits commercial use and changes to the model.
Who it works with
Organisations that want security testing run inside their own infrastructure, including air-gapped networks.
Key facts
- Size: about 504 billion parameters in total, roughly 40 billion active per token.
- Stored weights: 328 GB, down from 1.51 TB for the full-precision parent.
- Pruning: keeps 168 of the original 256 routed experts in each expert layer, selecting eight per token.
- Benchmark: Aikido reports average recall of 60.4% on 32 known vulnerabilities, against 65.6% for full-precision GLM-5.3. It measures targeted rediscovery, not blind discovery or exploit validation.
- Licence condition: larger companies that run a Model-as-a-Service business must pass a Z.AI security review before commercial use.
Summary written by TrustList from the business's own published information, Nov 2026.
Benchmarks & AI stats
| Benchmark | Official | Community avg |
|---|---|---|
| No benchmark scores yet. Be the first to add one. | ||
“Official” values are editor-approved and feed the ranking. “Community avg” is the mean of member submissions (shown for transparency; it never affects the ranking until an editor approves a value).
Sign in to add a benchmark score for this model.
Altar-1 launches
Trust Score
An earned signal from verification, reviews, awards, transparency and engagement - the vendor can't buy it.
Updated 9/22/2026
Request a demo or quote from Altar-1
Reviews
Questions & answers
No questions yet. Be the first to ask about Altar-1.
Altar-1 alternatives
Similar listings buyers compare against this one.
Microsoft AI's real-time speech-to-text model for voice agents, transcribing 60 languages with continuous language detection
Cantina Security's first open-weight security model, a post-trained GLM-5.3-Flash for vulnerability research and exploit checks
Microsoft AI's text-to-speech model for keeping one voice across 23 languages, with voice cloning from short audio
Open-source 2B decision model from AWS's Strands Agents project that picks options and returns scores with a reliability figure
JetBrains' open 12B mixture-of-experts coding model with 2.5B parameters active per token under Apache 2.0
Liquid AI's open-weight decision model, giving yes or no, choice or score answers from text and images
Perplexity's open-weight decision model that picks among given options for routing and checks in AI systems
More on TrustList
Everything here links back to the same verified catalogue. Pick your next stop.
- Write a reviewHelp other buyers decide
- CompaniesAgencies, consultancies and IT service providers, ranked by verified reviews.
- ProductsSoftware and SaaS with pricing, features, integrations and alternatives.
- ArticlesGuides, comparisons and research from the editorial desk and the community.
- AwardsAnnual recognition decided by verified reviews and an independent jury.
- LaunchesNew products and releases, voted up by the community every day.
- RequestsBuyers describe what they need; vendors respond directly.
- PeopleReviewers, authors and makers with public profiles.
- ComparePut up to four listings side by side before you shortlist.