LiquidAI Releases 350M-Parameter Multilingual Search Models "LFM2.5-Embedding-350M" and "ColBERT-350M"

On June 20, Liquid AI released two search-specialized models on Hugging Face — the first bidirectional models in the LFM family. Both have 350M parameters and support multilingual and cross-lingual retrieval across **11 languages, including Japanese**.

LFM2.5-Embedding-350M**: A dense bi-encoder (single-vector type). The lightest option, enabling the fastest index construction

LFM2.5-ColBERT-350M**: A late-interaction model (token-level vector type). Superior in accuracy and generalization

On the NanoBEIR multilingual benchmark, ColBERT-350M recorded an average NDCG@10 of 0.605, outperforming Qwen3-Embedding-0.6B (0.556) and Alibaba's gte-multilingual-base (0.528). On Japanese, ColBERT scored 0.614 and Embedding 0.575.

Inference latency on a MacBook Pro M4 Max is extremely fast at 7.3ms (p50) for query embedding. Quantized versions in GGUF format are also provided, making them usable as drop-in replacements for local RAG pipelines.

Practical use cases include multilingual product search in e-commerce, FAQ/support systems, on-chip semantic search, and enterprise knowledge assistants. The license is LFM Open License v1.0.

Bun 1.4: AI Rewrites the Codebase from Zig to Rust, Sparking Debate

Version 1.4 of the JavaScript runtime Bun has been released — and its contents are stirring up conversation: **the Bun team used AI to rewrite the existing Zig codebase in Rust**.

The effort has ignited fierce debate in the developer community. Opinions vary on the validity of large-scale AI-assisted code conversion: "Has the AI-driven rewrite received sufficient human review?" "Will architectural intent be lost?" "How do performance characteristics change?"

On the other hand, if this approach works reliably, it could have a major impact on modernizing legacy systems and migrating between programming languages. The fact that AI code generation is now being applied to whole-project rewrites — beyond individual functions and components — signals a new stage in software engineering.

Cutting LLM Pipeline Costs by 60% — TOON Format, Compressed Markdown, and Multi-Shot Examples

On Hacker News, an engineer who has built multiple LLM systems shared **practical cost-reduction techniques** that are drawing attention. Rather than switching models or fine-tuning prompts, they reportedly achieved substantial cost savings through data formats and instruction methods alone.

**1. Moving from JSON to the TOON format (about 30% fewer output tokens)**: JSON imposes significant syntactic overhead for LLMs. By adopting an alternative format called TOON, you can express the same information in fewer tokens.

**2. From full Markdown to compressed Markdown (about 50% fewer input tokens)**: For inter-agent communication and prompt context handoffs, use compressed Markdown and a short system prompt instead of full Markdown/HTML. Particularly effective for calls that carry over context from previous calls.

**3. From long Do/Don't lists to 2–3 multi-shot examples**: Counterintuitively, a few concrete examples improve output quality more consistently than detailed rule lists, while also cutting tokens. The longer a rule list grows, the more tokens it takes to cover real edge cases — making examples the more efficient choice.

The poster states that these changes alone delivered a 60% cost reduction — an approach worth trying before considering a switch to open-source models.

The June 2026 LLM Shutdown Rush — GPT-5 and o3 Snapshots and Claude 4 Deprecated in Rapid Succession

Qiita published **a roundup of shutdown and deprecation notices** aimed at teams using LLM APIs in production.

OpenAI has listed the deprecation of the GPT-5 and o3-series snapshots on its Deprecations page. The shutdown date is **July 23, 2026**, and the migration targets are the GPT-5.5/5.4 series. Meanwhile, Anthropic is also retiring the Claude 4 series.

The article organizes the action items you should check in your production code right now, making it a valuable reference for teams with little slack in their migration timelines. It once again highlights the importance of API version management and regularly monitoring deprecation schedules.

Cordium — A FOSS Sandbox Platform That Keeps Infrastructure Secrets Away from AI Agents

A Show HN post on Hacker News announced Cordium, a Kubernetes-based open-source sandbox platform. Its biggest difference from existing development environments and sandboxes such as GitHub Codespaces, E2B, and Daytona is that it provides **identity-based, credential-less access**.

Conventional sandboxes required injecting credentials — API keys, SSH private keys, database passwords, and so on — into the environment. Cordium keeps credentials behind an identity-aware proxy (Octelium) and injects them on the fly only for authorized users. This prevents AI agents inside the sandbox from directly accessing credentials.

With AI agents now entering widespread use, secret management for agents is an urgent challenge. Cordium's approach can be described as extending zero-trust thinking to AI agent execution environments. Being FOSS (free and open-source software) and self-hostable is also appealing for organizations that prioritize privacy and security.

---