What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Hirundo announced an $8 million seed round on June 9, 2025, to advance its machine-unlearning software: a system the company says can modify unwanted behavior in trained AI models without retraining them from scratch. Maverick Ventures Israel led the round.
What Hirundo raised the money for
Founded in 2023 by Ben Luria, Michael Leybovich, and Oded Shmueli, Hirundo develops enterprise software for machine unlearning. The company’s goal is to change unwanted information or behavior already learned by an AI model, rather than start over with a full retraining process. The funding announcement named SuperSeed, Alpha Intelligence Capital, Tachles VC, AI.FUND, and Plug and Play Tech Center as participants alongside lead investor Maverick Ventures Israel. Hirundo’s June 9, 2025 announcement
This is a business software service, not a consumer product. Hirundo’s current product site describes use before a model launches, in response to issues found in production, and as ongoing model hardening. It invites organizations to request a demo or apply for early access.
How Hirundo says machine unlearning works
Hirundo says its system identifies model behavior linked to problems such as hallucinations, bias, jailbreaks or prompt injections, toxic outputs, and memorized personal or confidential information. It then modifies the model to reduce the targeted behavior, without retraining from scratch, according to the company. CEO and co-founder Ben Luria compared the approach to “AI model ‘neurosurgery,’” describing a process that pinpoints and removes problematic knowledge or behavior within a model’s parameters. That is the company’s analogy, not independent technical validation.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
How it differs from filters and retraining
Hirundo positions its approach as an intervention in the model itself. By contrast, it describes guardrails and output filters as controls applied around the model’s outputs, and full retraining as potentially resource intensive. Those are the company’s distinctions; the available sources do not establish that filters or retraining are generally ineffective, or that changing a model this way preserves its useful capabilities in every context.
What the reported performance figures mean
Hirundo’s 2025 funding announcement reported the following results. The figures are company claims, not general performance guarantees:
| Claim in the announcement | Model named | What the evidence establishes |
|---|---|---|
| Up to 55% fewer hallucinations | Llama | Reported by Hirundo; the reviewed sources do not specify a detailed benchmark protocol or independent replication. |
| Up to 70% reduction in bias | DeepSeek-R1 | Reported by Hirundo; the reviewed sources do not specify a detailed benchmark protocol or independent replication. |
| 85% decrease in successful prompt injections | Llama | Reported by Hirundo; the reviewed sources do not specify a detailed benchmark protocol or independent replication. |
The announcement does not show that these percentages transfer to other models, datasets, or production environments. It also does not establish how much model utility is retained after intervention. A buyer evaluating the product would need model- and task-specific results, including the evaluation method and any impact on normal model behavior.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the funding announcement does—and does not—show
The round establishes that Hirundo announced seed financing led by Maverick Ventures Israel. It does not by itself validate the product’s performance or demonstrate commercial traction. The reviewed company materials do not state public pricing or offer a self-serve purchase flow; access is presented through a demo and early-access route.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Rank #3
For organizations assessing machine-unlearning tools, useful comparison points include whether a method changes model parameters or filters outputs externally, whether it works on an existing trained model or requires retraining, and what happens to utility on named benchmarks. Also ask which model families and deployment arrangements are supported, and whether claims have independent replication, production evidence, and disclosed cost or latency data. The available sources do not provide enough comparable independent evidence to rank Hirundo against other providers on those measures.
Quick Recap
Rank #4
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
