Tenkai Daily — April 29, 2026
Model Releases
- XiaomiMiMo-V2.5: Multimodal vision-language-audio agent — Multimodal agent supporting vision, language, audio, and video understanding with FP8 optimization. MIT license with eval-results for long-context benchmarks.
Open Source Releases
- claude-code v2.1.123: OAuth 401 retry loop fix — Fixes OAuth authentication failing with a 401 retry loop when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 is set, preventing infinite retry loops in restricted environments.
- claude-code v2.1.122: Bedrock service tier support — Adds ANTHROPIC_BEDROCK_SERVICE_TIER environment variable for AWS Bedrock deployments and improves PR URL session resumption.
- opencode v1.14.29: Session path & tool schema fixes — Sessions now preserve relative workspace paths, sanitizes Moonshot/Kimi tool schemas, and fixes MCP OAuth handling.
- jetson-examples 0.2.5 — One-line commands for running generative AI on NVIDIA Jetson devices, targeting edge deployment workflows.
- ACE-Step 1.5 Professional UI for Local AI Music Generation — Open-source UI for ACE-Step 1.5 enabling unlimited local music synthesis as a Suno alternative with advanced controls.
- Bitpoort — MCP server with 41 on-chain tools covering whale tracking, exchange flows, and cross-chain signals across Ethereum, Bitcoin, and Hyperliquid.
Research Worth Reading
- The Power of Power Law: Asymmetry Enables Compositional Reasoning — Power-law distributions in natural language actually help compositional reasoning by preserving long-tail skill frequencies, challenging the uniform-curation assumption.
- Architecture Determines Observability in Transformers — Defines ‘observability’ as linear readability of decision quality from mid-layer activations, showing architectural choices determine whether you can catch confident errors.
- minAction.net: Energy-First Neural Architecture Design — Evaluates 2,203 experiments across domains, showing explicit energy optimization yields Pareto improvements under computational constraints.
- Nautile-370M: Spectral Memory Meets Attention — 371M-parameter model interleaving linear-time spectral operators with transformers for efficient reasoning under tight budgets.
- Incompressible Knowledge Probes — Estimates undisclosed LLM parameter counts via factual capacity measurements, offering hardware-agnostic scale estimation.
AI Dev Tools
- goose v1.33.1: Windows CUDA build removal — Removes Windows CUDA build variant, standardizing CLI updates to simplify Windows deployment.
Today’s Synthesis
The convergence of XiaomiMiMo-V2.5 , jetson-examples 0.2.5 , and Nautile-370M points to a practical path for edge multimodal agents: MiMo-V2.5’s FP8 optimization + MIT license makes it deployable on Jetson hardware, while Nautile-370M’s spectral-attention hybrid demonstrates how to maintain reasoning quality under tight compute budgets. Engineers can now build vision-language-audio agents that run locally without cloud dependencies—useful for robotics, AR/VR, or privacy-sensitive applications where you need multimodal understanding but can’t afford the latency or cost of large model APIs.