--- title: Aikido releases Altar, an open-weight AI model for... description: Aikido Labs released Altar, an open-weight AI model built from GLM-5.3 for on-premises and air-gapped security testing. Using AWQ INT4 quantization and... canonical: https://daily.dev/posts/aikido-releases-altar-an-open-weight-ai-model-for-on-premises-security-testing-b38qp6lu4 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Aikido releases Altar, an open-weight AI model for on-premises security testing | daily.dev og:description: Aikido Labs released Altar, an open-weight AI model built from GLM-5.3 for on-premises and air-gapped security testing. Using AWQ INT4 quantization and... og:url: https://daily.dev/posts/aikido-releases-altar-an-open-weight-ai-model-for-on-premises-security-testing-b38qp6lu4 og:image: https://api.daily.dev/og/posts/b38QP6lU4.png og:image:alt: Aikido releases Altar, an open-weight AI model for on-premises security testing og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Aikido releases Altar, an open-weight AI model for on-premises security testing **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Aikido Labs released Altar, an open-weight AI model built from GLM-5.3 for on-premises and air-gapped security testing. Using AWQ INT4 quantization and Cerebras REAP-based expert pruning (removing 88 of 256 routed experts), the model shrank from 1.51 TB to 328 GB (78.2% reduction), fitting on a single 4×H200 node. On a 32-CVE internal benchmark, Altar retained 92% of the full-precision model's vulnerability recall. It powers Aikido Machine, an autonomous pentesting appliance that found a critical vulnerability shortly after production deployment. Weights, a fidelity study, and a pruning toolkit have been published, with plans to explore lower-bit formats like EXL3 and security-specific fine-tuning. ## Content Aikido Labs has released Altar-1, an open-weight AI model designed specifically for security use cases and built to run entirely inside customer infrastructure, including air-gapped environments. The starting point was GLM-5.3, a 1.51 TB full-precision model. To make it practical for on-premises deployment, the team applied AWQ INT4 quantization and then used Cerebras REAP-based expert pruning to remove 88 of the model's 256 routed experts. The result is a 328 GB model, roughly 78% smaller than the original and about 33% smaller than the quantized-but-unpruned version. It fits on a single 4×H200 node. On an internal benchmark of 32 known CVEs, Altar-1 averaged 60.4% recall and matched the quantized parent model's vulnerability coverage (23 of 32 CVEs), retaining about 92% of the full-precision model's coverage at a third of the storage cost. That's a reasonable trade-off for teams that can't or won't send data to external APIs. The model powers Aikido Machine, the company's autonomous pentesting appliance. It was deployed to production shortly after launch and reportedly found a critical vulnerability in its first run. Aikido has published the weights, a fidelity study comparing the pruned model against its parent, and a pruning toolkit. Next steps include experimenting with lower-bit formats like EXL3 and fine-tuning models specifically for security workflows rather than relying on general-purpose base models. It's early, but open-weight models with defensive security capabilities close to frontier performance have been notably absent. Altar-1 is a step toward filling that gap for teams with strict data sovereignty requirements. ## Questions this post answers ### How much smaller is Aikido's Altar model compared to the full-precision GLM-5.3 it's based on? Altar is 328 GB, a 78.2% reduction from GLM-5.3's full-precision size of 1.51 TB, and 32.8% smaller than the quantized parent model. This was achieved using AWQ INT4 quantization combined with Cerebras REAP-based expert pruning, which removed 88 of the model's 256 routed experts, allowing it to fit on a single 4×H200 node. _Teams evaluating self-hosted security models can follow model releases like this on daily.dev._ ### How much vulnerability detection accuracy does Altar lose from quantization and pruning compared to the full-precision model? Altar retains 92% of the full-precision model's vulnerability coverage despite being compressed to roughly a third of its storage size. On an internal benchmark of 32 known CVEs, it averaged 60.4% recall and matched the quantized parent model's coverage at 23 of 32 CVEs detected. _Developers weighing compression tradeoffs for security tooling can track benchmarks like this on daily.dev._ ### What is Aikido Machine and what model does it use for autonomous pentesting? Aikido Machine is Aikido Labs' autonomous pentesting appliance, powered by their Altar model, a compressed open-weight model derived from GLM-5.3. Shortly after being deployed to production, it found a critical vulnerability, offered as an early proof of concept for the system's effectiveness. _Anyone evaluating autonomous pentesting tools can follow real-world results like this on daily.dev._ ## Similar posts on daily.dev - [Aikido Machine: on-prem AI pentesting server](https://daily.dev/posts/aikido-machine-on-prem-ai-pentesting-server-izxogxhsz) · Aikido Security · 0 upvotes · 0 comments --- Tags: [#security](https://daily.dev/tags/security), [#machine-learning](https://daily.dev/tags/machine-learning), [#open-source](https://daily.dev/tags/open-source), [#ai-security](https://daily.dev/tags/ai-security) [View this post on daily.dev](https://daily.dev/posts/aikido-releases-altar-an-open-weight-ai-model-for-on-premises-security-testing-b38qp6lu4) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Aikido releases Altar, an open-weight AI model for on-premises security testing","url":"https://daily.dev/posts/aikido-releases-altar-an-open-weight-ai-model-for-on-premises-security-testing-b38qp6lu4","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/aikido-releases-altar-an-open-weight-ai-model-for-on-premises-security-testing-b38qp6lu4"},"datePublished":"2026-09-21T23:26:10.135Z","dateModified":"2026-09-22T01:52:15.665Z","description":"Aikido Labs released Altar, an open-weight AI model built from GLM-5.3 for on-premises and air-gapped security testing. Using AWQ INT4 quantization and...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/aikido-releases-altar-an-open-weight-ai-model-for-on-premises-security-testing-b38qp6lu4","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"security,machine-learning,open-source,ai-security","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Aikido releases Altar, an open-weight AI model for on-premises security testing"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/aikido-releases-altar-an-open-weight-ai-model-for-on-premises-security-testing-b38qp6lu4#faq","mainEntity":[{"@type":"Question","name":"How much smaller is Aikido's Altar model compared to the full-precision GLM-5.3 it's based on?","acceptedAnswer":{"@type":"Answer","text":"Altar is 328 GB, a 78.2% reduction from GLM-5.3's full-precision size of 1.51 TB, and 32.8% smaller than the quantized parent model. This was achieved using AWQ INT4 quantization combined with Cerebras REAP-based expert pruning, which removed 88 of the model's 256 routed experts, allowing it to fit on a single 4×H200 node. Teams evaluating self-hosted security models can follow model releases like this on daily.dev."}},{"@type":"Question","name":"How much vulnerability detection accuracy does Altar lose from quantization and pruning compared to the full-precision model?","acceptedAnswer":{"@type":"Answer","text":"Altar retains 92% of the full-precision model's vulnerability coverage despite being compressed to roughly a third of its storage size. On an internal benchmark of 32 known CVEs, it averaged 60.4% recall and matched the quantized parent model's coverage at 23 of 32 CVEs detected. Developers weighing compression tradeoffs for security tooling can track benchmarks like this on daily.dev."}},{"@type":"Question","name":"What is Aikido Machine and what model does it use for autonomous pentesting?","acceptedAnswer":{"@type":"Answer","text":"Aikido Machine is Aikido Labs' autonomous pentesting appliance, powered by their Altar model, a compressed open-weight model derived from GLM-5.3. Shortly after being deployed to production, it found a critical vulnerability, offered as an early proof of concept for the system's effectiveness. Anyone evaluating autonomous pentesting tools can follow real-world results like this on daily.dev."}}]} ```