BuzzRAG AI Desk — 2026-10-22

IBM has released Granite Speech 5.0 2B, a speech recognition and translation model distributed under the Apache 2.0 license, according to the release report and a corroborating Hugging Face blog post. The reported Open ASR Leaderboard result is second place, with 5.68% average word error rate on English. IBM also updated its Turbo CTC and Granite 4.1 2B Plus models, though the supplied report does not detail those changes.

The license and model size make this release relevant to teams evaluating systems they can adapt and run in their own environments. But a leaderboard position is only one measure: the cited figure is for English, and the available information does not specify the model’s results across other languages, test conditions, or translation quality. Those details will matter for organizations weighing real multilingual workloads rather than a single aggregate score. The next useful comparison is performance across languages and deployment settings, alongside the cost of running the 2-billion-parameter model.

A report from GTC Berlin says Revolut and other major European financial firms are adopting open-weight models supported by NVIDIA infrastructure. The supplied coverage offers few specifics: it does not name the participating firms beyond Revolut, identify the models, or describe whether the systems are in production, pilots, or procurement discussions. The headline’s claim that banks are “ditching” closed AI therefore goes further than the evidence provided.

For regulated financial institutions, open weights can offer more control over where models run and how they are adapted, but they do not automatically make a system fully open or remove the need for security, governance, and auditing. Infrastructure dependence also remains a consideration when deployment is tied to a particular hardware stack. The consequential questions are what tasks these firms are assigning to the models, how they validate outputs, and whether reported adoption means sustained operational use. Without those details, this is an indication of interest—not evidence of a sector-wide replacement of closed systems.

Researchers from Perplexity and NVIDIA have introduced JOLT, a post-training method that uses one model in both teacher and student roles. The reported result is that JOLT matched GRPO accuracy on the MATH benchmark while processing 6.5 times fewer training tokens. That is a potentially useful efficiency result, but the supplied summary does not give model sizes, the precise evaluation setup, or results beyond MATH.

The approach targets a costly part of model development: reinforcement-learning post-training, where models are optimized using feedback rather than only next-token prediction. Reusing a model across teaching and learning roles may reduce the amount of generated training data needed, though token count alone does not establish lower total compute or faster training. The comparison will be more informative when researchers can assess wall-clock time, hardware requirements, performance across other reasoning tasks, and whether the result holds at different scales. For now, JOLT offers a specific benchmark claim, not evidence that reinforcement-learning costs have broadly fallen by the same factor.

Reports say attackers tried to exploit vulnerabilities described as “PHP2Shell” before fixes were available, with Amazon Web Services logging attempts four days before the vulnerabilities were patched. The flaws are now described as being broadly exploited across the internet, and the story has been corroborated by several outlets. The supplied information does not include vulnerability identifiers, affected versions, or technical details, so the scope and attack path cannot be assessed from this account alone.

The reported gap between attempted exploitation and patch availability underscores a familiar security problem: public-facing software can be targeted while maintainers and operators are still responding. For organizations using affected PHP components, the immediate priority is to identify exposure, apply the relevant vendor fixes, and review logs for signs of compromise. The wider AI-desk relevance is indirect but important: AI services depend on ordinary web infrastructure, and an exposed application can put data and connected systems at risk regardless of how capable the models are. The key follow-up is whether incident responders publish clearer indicators and guidance on affected deployments.

Meta reportedly beat revenue estimates for the third quarter and raised its outlook, while signaling that investment in AI is increasing. The available summary does not provide the revenue figures, revised guidance, or a breakdown of AI-related spending, so the scale of the change cannot be quantified here. The result is nonetheless another sign that large technology companies are tying their plans to continued investment in AI infrastructure and products.

Higher spending can support the compute, data-center capacity, and engineering needed to train and serve models, but spending alone is not proof of better products or returns. Investors and customers will need to distinguish capital devoted to AI systems from broader infrastructure costs, and then see whether those investments improve engagement, advertising performance, or other measurable business outcomes. Meta’s guidance will be more revealing when paired with details on where the spending is going and how quickly it translates into deployed capabilities. For the wider field, the earnings report is a reminder that the race to scale is also a test of whether AI investment can justify its cost.