Aleph Alpha releases Kolibri, a 78B sovereign open-weight model under Apache 2.0
Aleph Alpha released Kolibri on 2026-10-03: a mixture-of-experts transformer with 78.1B total and 3.46B active parameters (384 experts, 6 active, 50 layers, up to 1M-token context), trained on about 24T tokens (62% English, 21% German, 14% code) in Germany and Finland. Weights are on Hugging Face under Apache 2.0. The company reports AIME 2025 at 96.9% and HumanEval+ at 92.7% and says it matches models with up to 4x the active parameters; those are vendor-reported figures, not independently measured. It is aimed at regulated sectors such as public administration and aerospace, and is notable as a permissively licensed European open-weights model with very low active-parameter cost.
Sources & depth
Microsoft's ThinkingBox finds agents that pass once often fail on repeat
Microsoft released ThinkingBox, an open benchmark of 507 stateful business tasks, each run 20 times against a freshly reset database. It grades the backend state the agent leaves behind rather than its tool calls or final message. Single-attempt success is a poor predictor of reliability: Claude Opus 5.5 passes 67.16% of tasks once but only 47.53% on all 20 attempts, and Kimi-K3 solves 93.89% at least once but completes just 13.41% consistently. The authors attribute 79.9% of failures to tool handling rather than reasoning. Dataset (CDLA-Permissive-2.0), framework (MIT) and harness are public on Hugging Face via OpenEnv.
Sources & depth
Anthropic publishes prompting guidance for Opus 5.5 in Claude and Claude Code
A new guide says Opus 5.5 runs longer multi-step tasks with less intervention, reasons before each reply by default, and reports what it did and what it needs next. Recommendations: give tasks explicit finish lines, drop "think carefully" prompts as redundant, add stopping rules to CLAUDE.md, split large migrations across subagents with verification, and keep task checklists in files so they survive context summarization. It also points to /fast in Claude Code (research preview, higher token cost). This is usage guidance, not a new release.
Sources & depth
Reports: OpenAI safety leader resigns, citing a "broken" company culture
The Guardian and The Atlantic both ran pieces on 2026-10-03 about a senior OpenAI safety figure quitting and warning the company's culture is broken. Both outlets block automated fetching, so the details (who, what was alleged, OpenAI's response) are unverified here; read the sources directly.
Sources & depth