Liquid AI has launched two new bidirectional encoders named LFM2.5-Encoder-230M and LFM2.5-Encoder-350M. Both models handle a long context window of 8,192 tokens, which is about 13 to 15 pages of text. These encoders are built on the LFM2 hybrid backbone and start from decoder backbones LFM2.5-230M and LFM2.5-350M.
Unlike causal models, these encoders use bidirectional attention masks. They also replace causal short convolutions with symmetric center padding, making them non-causal. The training uses a masked language modeling objective with a 30% mask rate. Training happens in two steps: first at 1,024 tokens on a large web corpus, then extending to 8,192 tokens for longer contexts.
Both models have a hidden size of 1024 and support 15 different languages. The license for these encoders is the LFM Open License v1.0, allowing open use. Liquid AI ran benchmarks on 14 models across 17 tasks from GLUE, SuperGLUE, and multilingual classification tests.
The LFM2.5-Encoder-350M scored a mean of 81.02 (±1.00) across 17 tasks, ranking fourth overall. The smaller 230M version scored 79.29 (±1.02), landing in sixth place. Both beat Liquid AI’s own retrieval-focused models, LFM2.5-ColBERT-350M and LFM2.5-Embedding-350M. These encoders are designed for edge devices, regulated systems, and pipelines that handle high volumes of data.
They can be loaded using transformers with auto_map, but fine-tuning is required to get good general-purpose representations. This makes them flexible for a range of real-world applications, especially where long context and multi-language support matter.
Anthropic’s Opus 5 and Industry Reactions
Anthropic recently released Opus 5, a powerful new AI model. It performs about the same or slightly better than Anthropic’s Fable model on coding tasks. Opus 5 also beats Anthropic’s previous version Opus 4.8 and OpenAI’s GPT-5.6-Sol in multiple benchmarks. However, it is not specialized for cybersecurity and falls short of the Mythos 5 model in vulnerability detection.
Opus 5’s pricing is notable. It costs $5 per million input tokens and $25 per million output tokens. This is more expensive on output tokens than the Chinese open-weight model Kimi K3, which charges $15 per million output tokens. Kimi K3 was developed by Moonshot, a Chinese AI lab.
Industry voices have criticized Anthropic’s stance on open models. Bill Gurley pointed out that Anthropic’s reluctance to support open weight AI reflects their corporate strategy. Peter Steinberger from OpenAI noted Anthropic has stayed silent on open weight licensing. Kai-Fu Lee called attention to who did not sign the Open Weight letter, implying Anthropic’s absence is telling.
David Sacks summed up the industry’s mood: “The entire tech industry (save for Anthropic) has come out in favor of open source AI.” Clem Delangue, CEO of Hugging Face, called for “an unprecedented response” after a recent security breach involving OpenAI models hosted on Hugging Face.
Microsoft Steps Up AI Cybersecurity Efforts
Microsoft launched Project Perception on July 27, 2026. It’s a security platform using specialized AI agents named Red, Blue, and Green to detect and defend against cyberattacks. One key tool, MAI-Cyber-1-Flash, performs about 95% of the work done by Microsoft’s MDASH vulnerability-finding system.
When combined with GPT-5.4, this system scored 95.95% on the CyberGym benchmark. This shows strong AI-driven performance in finding and exploiting security vulnerabilities. Microsoft plans to offer MAI-Cyber-1-Flash through its Azure AI Foundry, making it widely accessible to enterprises.
David Weston, Microsoft’s corporate vice president of AI security, said, “We’re not going to let the attackers have all the productivity increase.” This highlights the growing role of AI in cybersecurity defense.
Meanwhile, security breaches involving OpenAI models have stirred debate about AI risks. Hugging Face’s CEO Clem Delangue urged swift action after the breach on his platform. This incident has pushed the industry to focus more on AI security and open source transparency.
As AI models grow more powerful, the race to balance openness, performance, and security heats up. New releases like Liquid AI’s encoders and Microsoft’s AI security tools show innovation moving forward on multiple fronts.
Based on
- Liquid AI Releases LFM2.5-Encoder-230M and LFM2.5-Encoder-350M: Bidirectional Encoders That Stay Fast at 8K Context on CPU — marktechpost.com
- Anthropic gets heat for being the only major AI lab not supporting open models | Business Insider Africa — africa.businessinsider.com
- Anthropic’s Opus 5 is about token efficiency, not a capability leap – Ars Technica — arstechnica.com
- Hugging Face CEO shares his demands of OpenAI after ‘rogue’ agent hack: ‘It deserves an unprecedented response’ | Business Insider Africa — africa.businessinsider.com
- Microsoft Project Perception launches AI agents, specialized model for cybersecurity — axios.com
