Daily Executive Summary
This run adds six events and updates four existing ones. The recall audit recovered Kolibri, Naive-N0.5-Flash and DwarfStar ds4. Strands Decider makes agent routing reproducible, Microsoft released low-latency speech models, and Index-Translate opened constrained-translation weights. Claude Code and the MCP SDK tighten execution and transport boundaries, while Copilot code review is now API-orchestrated. UI-Venus-2's artifacts are public, but its license is unsettled.
Models
Strands Decider 2B (TRIAL)
Strands opened a 2B decision model that selects among typed actions in one forward pass without generating text. A decision model only answers “which action?”; the hard parts are calibration, abstention and the consequence of a wrong choice, not fluent prose. The weights, training code, data and evaluation records support a one-day trial on real routing samples.
MAI Speech (TRIAL)
Microsoft released streaming transcription for 60 languages and two speech-synthesis models covering 23 languages. Streaming transcription returns partial text before an utterance ends; the difficult part is handling revisions, interruption and rollback correctly. Test mixed Chinese-English calls for first-token latency, error rate, interruption recovery, naturalness and per-turn cost.
Index-Translate (TRIAL)
IndexTeam opened a translation MoE with 35B total and 3B active parameters, supporting 150 languages plus terminology, style and JSON constraints. “3B active” means each request uses only part of the expert capacity to reduce inference cost; routing overhead and memory still need measurement. It is a preview, so run one day of terminology and low-resource samples with the public weights.
Kolibri (TRIAL)
Aleph Alpha opened the German-English Kolibri model: 78B total parameters, 3.46B active, reasoning levels, tool calling and up to 1M tokens. Its weights, vLLM plugin and training report support German enterprise RAG and long-document trials. The vendor still recommends at most 262K for complex tasks, and author benchmarks need reproduction.
Naive-N0.5-Flash (TRIAL)
NaiveAI opened a sparse model with 309B total, 15.5B active and native 1M context. It uses no full-attention layers, but the weights alone occupy roughly 315 GB and require an FP8 NVIDIA cluster. This is a targeted trial for teams with suitable hardware, not a default deployment choice.
Agent & AI Engineering
MCP SDK 2.3 (TRIAL, updated)
The TypeScript and Python SDKs tighten same-origin redirects, token audiences, connection lifecycles and SSE payload limits. The safer defaults are useful, but older implementations that reuse servers or transports can break. Cover cross-origin redirects, oversized events and reconnection before rollout.
UI-Venus-2 (TRIAL, rating changed)
The due review confirmed that the 9B weights, framework, training and evaluation tools are public, moving the item from WATCH to TRIAL. The model card explicitly says the weight license is still pending. Trial only in isolation and wait before commercial use.
Open Source
Strands Decider, Index-Translate and UI-Venus-2 all publish usable artifacts, but with different openness. The first two use Apache-2.0; UI-Venus-2's weight license remains unresolved. License status belongs in the same adoption gate as model quality.
Research
The verifier-audit 30-day review found no new independent adoption signal. Its MIT code and per-sample records remain reproducible, so the rating stays TRIAL. It is better treated as a method for checking an internal reward pipeline than as a mature tool.
Developer Tools
Claude Code 2.1.288 (ADOPT, updated)
The release fixes duplicate tool calls after oversized or unparseable MCP output and fails closed when hook state cannot be serialized. It also catches destructive removal hidden inside bash -c and improves recovery after auto-compaction. Regress dangerous shell, hooks, oversized output and recovery before upgrading.
Copilot Code Review (TRIAL, updated)
GitHub now lets REST and GraphQL clients request Copilot code review and set effort per request; Balanced is the default GA tier. This makes CI integration practical, but teams should compare Balanced and High on the same PR set for latency, cost and useful-finding rate.
Infrastructure
DwarfStar ds4 (TRIAL)
The recall audit recovered ds4, a C inference stack specialized for DeepSeek V4, GLM 5.x and Qwen3.8. It provides SSD streaming, persistent KV caches, local APIs and a native coding agent, with more than 23,000 GitHub stars. The project remains a fast-changing beta and needs full performance and parity testing on target hardware.
Business & Policy
No new policy event changed model access, API cost, chip supply or the open ecosystem this cycle.
Trend Signals
No evidence-supported new trend emerged. MCP SDK 2.3 strengthens the existing enterprise security and governance signal. Copilot API orchestration and Claude Code long-session recovery extend the coding-agent runtime trend, while Claude Code's fail-closed behavior reinforces the separation of execution capability from policy control. Status and confidence remain unchanged.
Tech Radar
- New Strands Decider (foundation-model / TRIAL): Public weights and complete experiment artifacts make agent routing independently testable.
- New MAI Speech (foundation-model / TRIAL): Callable APIs and published pricing support an end-to-end speech-latency trial.
- New Index-Translate (foundation-model / TRIAL): Open weights enable constrained-translation tests, but preview status and vendor benchmarks do not support production use.
- New Kolibri (foundation-model / TRIAL): German-English open weights and a serving plugin make enterprise RAG and tool use directly testable.
- New Naive-N0.5-Flash (foundation-model / TRIAL): The native-1M sparse design is testable, but roughly 315 GB of weights limits its audience.
- New DwarfStar ds4 (infrastructure / TRIAL): Strong adoption and complete local-inference features come with fast-changing beta stability.
- UI-Venus-2 (ai-agent / TRIAL): Public weights and framework are verified, moving it up from WATCH; the license still blocks commercial adoption.
- MCP SDK 2.3 (ai-agent / TRIAL): Stronger security defaults come with transport-lifecycle compatibility work.
- Claude Code 2.1.288 (developer-tools / ADOPT): The release reduces duplicate side effects and tightens dangerous-command and hook boundaries.
- Copilot Code Review (developer-tools / TRIAL): API access makes CI integration possible, while quality and cost still need testing.
Current 30-day radar: ADOPT 9 / TRIAL 113 / WATCH 76 / IGNORE 0. See the site Insights page for the full current radar.
Worth Trying
Prioritize two one-day experiments: compare Strands Decider with the current general model on real routing samples; German-language teams should test Kolibri on RAG and tool calling. Teams with high-memory workstations or FP8 clusters can separately evaluate ds4 or Naive-N0.5-Flash.