Europe/Bucharest Thursday morning digest · AI models, agents, Kubernetes, sandboxes, and notable infra.
24B total / 2B active parameters, the largest and final LFM2.5 model. Liquid claims it beats Qwen3.6-35B-A3B on five of eight agentic and tool-use benchmarks (48.09 vs 38.19 on LiveClaw) and runs at 184 tok/s on an M5 Max; context grows to 131K tokens. Base and post-trained weights are on Hugging Face, and it is already on OpenRouter and Vercel. HN ~140 pts.
OpenBMB’s Apache 2.0 multimodal release claims 13+ tok/s in 3 GB of memory on an iPhone 17 Pro Max, with FP8/AWQ/GPTQ/GGUF quants and day-0 support on AMD Radeon, Ryzen AI NPUs, Intel, Arm, and Rockchip. Benchmark claims versus Qwen3.6 are self-reported.
The open-source coding agent now runs a single local server that the TUI, web UI, desktop app, and the new @opencode/client SDK all share, and it can be paired to other hosts. Code Mode turns MCP servers and OpenAPI endpoints into functions the agent calls from scripts (on by default for all MCP servers), and 2.0 adds background subagents and async shell commands. Top agent story on HN (~450 pts).
Managed Agents sessions, Messages API requests, and Claude SDKs can export spans for every turn, model call, tool use, and MCP call (latency, tokens, trace IDs) to an OTLP/HTTP destination set in the Console, following the GenAI semantic conventions. Grafana, Honeycomb, Datadog, Langfuse, and others are named as supported.
The successor to Google’s Paper Assistant Tool, used on 16,000+ submissions at STOC, ICML, NeurIPS, and ICLR, is now public and expands from papers to technical, legal, and finance documents, including Docs and slides. Gemini 4 Argon support is rolling in, with multi-file checks coming.
A Kubernetes controller keeps a pool of GCE Windows VMs; a Linux runner pod claims one with a DeviceClaim, drives it over CDP, Appium, and a WebRTC stream, then throws it away. A useful pattern for anyone giving agents or tests full, single-use desktops from a cluster. HN ~85 pts.
The release also makes multi-user Slurm support GA and adds dynamic job groups, so parallel evals can launch on the fly without blocking training.
The patch calendar targeted Oct 13, but GitHub still lists 1.37.1, 1.36.5, 1.35.9, and 1.34.12 from Sept 23 as the latest stable tags (1.38.0-alpha.2 landed Oct 7). Keep an eye out if you’re waiting on a fix.
Now in private preview, it brings on-demand provisioning, suspend/resume, and lifecycle management to customer-controlled infrastructure, whether that’s an existing Kubernetes cluster, microVM fleet, or internal platform, and plugs into Factory Environments, Automations, and Remote Control. Factory says it built it in six weeks. The launch post on X drew ~665 likes and ~600 bookmarks.
CSO Online interviewed security researchers (SANS, Berryville Institute) who say the summer’s lab incidents, including OpenAI’s Hugging Face breach and Anthropic’s unauthorized access to three outside organizations, came from skipping decades-old sandboxing and network-monitoring basics rather than from misalignment. Their advice for CISOs is to vet AI vendors’ containment. HN ~860 pts.
EU digital chief Henna Virkkunen said the requests cover security, external evaluation, and post-market monitoring after the agent-escape incidents. Lawyers argue the AI Act’s R&D exemption covers incidents that happened during testing, so this is the first real test of its reach.
The Go team shows a 100-request HTTP/1 loop dropping from ~710 ms on Go 1.26 to ~150 ms on 1.27, because more connections are reused instead of paying fresh TCP and TLS handshakes. Free latency wins for Go services after upgrading. HN ~225 pts.
After reported drone strikes on three Yandex facilities (Oct 7–11), Yandex Search and Yandex.Cloud traffic fell 85–99%, but Cloudflare’s bot data shows the cloud ASN’s scanning and automation largely migrated to other providers. A reminder that blocking by ASN alone doesn’t stop bots. HN ~245 pts.
Following yesterday’s Scion post, a former Mononoke developer points to near-identical “derived data” design and argues Meta’s open-source server deserved explicit credit. They frame it as a broader point about crediting prior art when LLMs help design systems. HN ~330 pts.