AI Daily | Kolibri Open-Source MoE Model, Perplexity Visualize Feature, and Meta RankEvolve Multi-Agent Framework

AI Daily | Kolibri Open-Source MoE Model, Perplexity Visualize Feature, and Meta RankEvolve Multi-Agent Framework

Model Releases & Updates

Kolibri — Aleph Alpha — Aleph Alpha

• Bottom line：Aleph Alpha has released Kolibri, a 78.1B parameter open-weight German-English bilingual MoE model featuring a 1M token context window under the Apache 2.0 license.
• Architecture：Employs a Mixture-of-Experts architecture with 78.1B total parameters and 3.46B active parameters per token.
• Context & Openness：Supports up to 1 million tokens of context and is fully open-weighted under the Apache 2.0 license.
• Source：Tech Report
• Source：Hugging Face

Sopro V2 Turbo 2610 — Samuel Vitorino — Samuel Vitorino / LocalLLaMA

• Bottom line：An interim update to the lightweight Sopro text-to-speech model focusing on cleaner voice cloning and reduced roughness while maintaining a 120M parameter size.
• Performance：Achieves ~300ms to first audio on laptop CPUs with a 120M parameter architecture.
• Languages：Supports English, European Portuguese, French, and German under an Apache 2.0 license.
• Source：GitHub Repository
• Source：Hugging Face Weights

Product Releases & Updates

Visualize Feature in Perplexity Computer — Perplexity — Perplexity

• Bottom line：Perplexity introduced a new Visualize capability in Computer, generating inline interactive components and animations directly within user chat responses.
• Functionality：Users can prompt queries starting with ‘Visualize’ to receive dynamic educational widgets and animations.
• Recommendation：Recommended for use under Standard or High effort modes for optimal visual output.
• Source：

  View post on X by @AravSrinivas

  在 X 上查看 @AravSrinivas 的貼文 ↗

Extract v2.5 — LlamaIndex — LlamaIndex

• Bottom line：LlamaIndex launched Extract v2.5, a series of frontier agents tuned for precise multi-page document extraction and complex tabular parsing.
• Performance：Outperforms Opus 5.5 and GPT-6 Sol on complex long-list and multi-page table extractions while being 30% to 4x cheaper.
• Accuracy Gains：Achieved a jump from 86.1% to 95.5% on complex extraction tasks using the agentic tier.
• Source：Blog Announcement

Industry News

AWS Ends Government NDA Usage for Data Centers — Amazon Web Services — Amazon Web Services

• Bottom line：AWS CEO Matt Garman announced that the company has stopped using non-disclosure agreements when partnering with government agencies on data center construction.
• Policy Shift：Eliminated NDAs for government-linked data center projects to increase transparency amid public scrutiny.
• Resource Defense：Garman noted that direct data center water usage accounts for only 0.5% of US industrial consumption and highlighted $1 billion in community investments.
• Source：TechCrunch Coverage

Cloudflare Launches Next-Gen Git Platform Challenge — Cloudflare — Cloudflare

• Bottom line：Cloudflare announced an open beta for Artifacts and launched a developer challenge to build multi-agent Git platforms using Workers and Artifacts.
• Competition Details：Invites developers to build agent-era Git platforms requiring multi-agent concurrency, running until October 14, 2026.
• Prizes：Offers $25,000 in Cloudflare credits for the first-place winner.
• Source：Cloudflare Blog

Research Papers

RankEvolve: Multi-Agent Auto-Research Framework — Meta — Meta

• Bottom line：Meta introduced RankEvolve, a multi-agent framework where coding models cross-review and refine research code modifications under compilation protocols.
• Methodology：Leverages models like Claude Code and Codex as independent nodes to mutually review and patch code changes.
• Results：Boosted task execution accuracy from 45.8% to 62.5% and improved recommendation model benchmarks.
• Source：

  View post on X by @dair_ai

  在 X 上查看 @dair_ai 的貼文 ↗

The Sharpening Tax in RL Post-Training — Meta Superintelligence Labs — Meta Superintelligence Labs

• Bottom line：Researchers discovered that reinforcement learning post-training improves pass@1 accuracy but imposes a ‘Sharpening Tax’ that limits test-time scaling performance.
• Finding：Base models with lightweight inference frameworks frequently outperform RL-tuned models when given sufficient sampling budgets.
• Proposed Solution：Introduced PTGS to estimate prompt difficulty and set sampling temperatures to mitigate the scaling tax.
• Source：

  View post on X by @omarsar0

  在 X 上查看 @omarsar0 的貼文 ↗

ThinkingBox Agent Sandbox & Benchmark — Microsoft & Hugging Face — Microsoft / Hugging Face

• Bottom line：Microsoft and Hugging Face released ThinkingBox, an agent sandbox and benchmark designed to evaluate stateful business workflows using database end-states.
• Scope：Covers 507 stateful business workflows with 20 repeated evaluations per task.
• Execution：Uses executable evaluation based on final database states and side effects, accessible via OpenEnv.
• Source：Hugging Face Blog

Other Highlights

Advocating Default Hard Budget Caps for Metered AI Services — Simon Willison — Simon Willison

• Bottom line：Developer Simon Willison published an essay calling for metered AI services and APIs to provide mandatory default hard budget caps to prevent runaway agent bills.
• Core Argument：Usage-based services should immediately cut off and return errors upon hitting limits rather than relying solely on warning emails.
• Ecosystem Impact：Autonomous coding agents lower the barrier to deploying paid code, escalating the risk of accidental large invoices.
• Source：Blog Post