A Belgian Firm Slimmed a Chinese LLM by 78% Into a Security Auditor Banks Dare to Use
Banks want AI to hunt code vulnerabilities, but the code can't leave the building. The Belgian answer: chop 78% off a Chinese open-weight model and move it into the basement.
Three key facts
On September 21, Belgian cybersecurity firm Aikido released Altar, an open-weight security model compressed and customized from Z.AI's GLM-5.3, purpose-built for local deployment — organizations can inspect security-sensitive code in their own data centers without sending a single line to an external inference provider. The compression combines quantization with expert pruning: 168 of GLM-5.3's 256 routed experts retained, storage cut from 1,506.7GB (full-precision GLM-5.3) to 328GB (-78.2%), servable on a single node with four H200 GPUs. Pruning was calibrated on internal penetration-testing traces and multilingual text — not customer data.
The benchmark disclosure is unusually honest: the internal test covers 32 known vulnerabilities across 30 code repositories, three runs per case. Altar averaged 60.4% recall and rediscovered 23 of 32 at least once; the quantized GLM-5.3 baseline averaged 61.5% with the same 23; the full-precision parent hit 65.6% and 25. Compression cost about one percentage point of recall while keeping 92% of the parent's vulnerability coverage. Aikido also volunteered the benchmark's limits: it measures known-CVE rediscovery inside a pipeline — not blind whole-codebase discovery, exploit validation, or fix quality.
The commercial loop is already closed: Altar's weights are publicly downloadable, and the model will ship inside Aikido's own products — whose customer list includes Belgian bank Belfius. Financial institutions are exactly the buyers for whom "code never leaves the intranet" is a hard requirement.
Wang Dou's Take
The most delicious part of this story is the supply-chain reversal. For years the "security" narrative treated Chinese open-source models as a risk source; now a European security firm takes GLM-5.3 as its parent stock, prunes it, quantizes it, calibrates it, and sells the result as bank-grade product to Belfius. That's the beauty of open weights: a model's birthplace doesn't determine its use — whoever holds the weights gets to reshape it. Z.AI doesn't collect a cent in licensing fees, yet the GLM line has become bedrock under European financial security infrastructure. Open source never pays out on invoices; it pays out in ecosystem position.
Then look at how Aikido reported its numbers — a rare species in this industry: not just its own 60.4%, but the quantized baseline's 61.5% and the parent's 65.6% on the same table, plus an unprompted declaration of what its benchmark does not measure. Compare that with vendor launches that report only their strongest event, and drawing your own measurement boundaries in public is exactly what a security company should look like. You are selling "trustworthy"; hiding benchmark edges is like writing "probably effective" on the product label.
The practical takeaway for enterprises fits in one line: four H200s plus 328GB of weights is the complete price list for "AI security auditing that never leaves the intranet." This multiple-choice question used to have two options — hand your code to a cloud API, or skip AI entirely. Now there's a third: bring the open model home and raise it yourself. Code-vault keepers at banks, defense firms and pharma companies should be adding a line to this year's budget sheet.
Source: Aikido Blog · SuperpowerDaily · Investing.com