Daily Executive Summary

  • China's AI export catalogue remains a draft, but temporary registration is live: 23 AI technologies are proposed for addition and comments run through November 15. Teams should inventory cross-border transfers now without misreporting the draft as a blanket ban.
  • Beam moves from preview to Apache-2.0 open weights: the 501B/23B-active model is testable through several runtimes or a hosted endpoint; an 8×H200 reference setup at about $40 per hour and vendor-run benchmarks remain key constraints.
  • NVIDIA releases Nemotron 3.5 multimodal weights: Super VL offers 1M context and Bolt has a 24GB artifact; OpenMDW terms, hardware compatibility and vendor-reported results still need validation.
  • Granite Switch 4.1 uses one base for several adapters: Apache-2.0 8B/30B models put chat, decision, guardian and ASR behind one endpoint, making memory and service-consolidation claims testable.
  • llama.cpp 0.7.0 improves MoE inference and changes interfaces: expert caching may speed decoding, while the port, chat response shape and RPC protocol all require regression testing.

Models

  • Beam (updated / TRIAL): BF16, FP8 and NVFP4 weights, a model card and several runtimes are available. Sparse MoE activation does not make deployment cheap; the reference hosted setup still uses eight H200s. Release
  • Nemotron 3.5 (new / TRIAL): Super VL targets long-context multimodal agents, while Bolt provides a smaller 24GB quantized path. Most results come from NVIDIA or partners. Announcement
  • Granite Switch 4.1 (new / TRIAL): composable adapters share a base model and KV cache. ASR is a cascade, not jointly trained multimodality. Release
  • Liquid d1-ultra (updated / TRIAL): the hosted API returns fixed-choice probabilities at $0.042 per million input tokens with no output-token charge; its 112-task results are vendor-run. Announcement
  • Timestral (new / TRIAL): 5M–209M open weights lower the cost of time-series baselines; leaderboard advantages need independent reproduction. Announcement

Agent & AI Engineering

  • Claude Code 2.1.301 (updated / ADOPT): fixes workflows that skipped a step yet reported success, excessive unattended retries and lost worktrees. Stable release
  • Codex CLI 0.163.0 (updated / ADOPT): runs read-only tools in parallel, lets steering interrupt long tool calls, and continues sandbox and resume fixes. Stable release

Open Source

Beam, Nemotron 3.5, Granite Switch 4.1 and llama.cpp provide inspectable artifacts, but licenses, hardware needs and migration risks differ. “Open weights” does not replace adoption testing.

Research

No independent research paper was added to the event library in this run. Google Willow magic-state cultivation is an important quantum-computing experiment, but it has not demonstrated practical scale and did not displace the five daily highlights.

Developer Tools

Stable Claude Code and Codex releases continue to turn long-task reliability into testable engineering boundaries: failure state, resume, worktrees and steering belong in upgrade regression suites.

Infrastructure

  • llama.cpp 0.7.0 (new / TRIAL): MoE expert caching keeps frequently used experts on GPU. Gains depend on model and memory allocation, and the 2.2x figure is project-reported. Release

Business & Policy

  • China AI technology export rules (new / WATCH): catalogue additions remain under consultation, while temporary pre-export registration is effective. The official Q&A states important boundaries for public-domain open source, APIs, SaaS and lightweight edge delivery. Announcement · Q&A

Trend Signals

  • Trend #3: Codex and Claude Code further strengthen parallel execution, steering and failure recovery; status remains established.
  • Trend #5: Beam, Nemotron and Granite add local-weight paths, supporting the enterprise self-hosted execution-plane candidate.
  • Trend #7: truthful workflow failure states, sandboxes and permission recovery further support separating execution from consequence control.
  • No sufficiently evidenced new trend emerged in this cycle.

Tech Radar

  • New · Nemotron 3.5 (foundation-model / TRIAL): weights and hosted paths are available; validate license, hardware and a real multimodal task.
  • Beam (foundation-model / TRIAL): moves from limited preview to Apache-2.0 open weights, upgrading from WATCH to TRIAL.
  • New · Granite Switch 4.1 (foundation-model / TRIAL): shared-base adapters make service-consolidation gains testable.
  • New · llama.cpp 0.7.0 (infrastructure / TRIAL): MoE optimization is useful, but API migration needs regression testing.
  • New · AI export registration (business-policy / WATCH): temporary registration is live, while the final catalogue remains unsettled.
  • Claude Code 2.1.301 (developer-tools / ADOPT): fixes false workflow success and unattended recovery risks.
  • Codex CLI 0.163.0 (developer-tools / ADOPT): parallel reads and in-flight steering improve daily agent workflows.
  • Liquid d1-ultra (foundation-model / TRIAL): a low-cost hosted path is testable against local d1 on one labeled set.

Current radar: ADOPT 9 / TRIAL 123 / WATCH 55 / IGNORE 7. See the site Insights page for the full current radar.

Worth Trying

  • Compare Nemotron 3.5 with the current model on one multimodal tool task and hosted cost.
  • Run one real repository task on Beam, recording success, long-context retention and cost per success.
  • Before upgrading llama.cpp, regress ports, response parsing, prompt cache and RPC.

Sources

Coverage note: all 36 fixed monitors were fetched and reviewed; active search also covered open source, small teams, Chinese policy sources and exceptional general technology. NVIDIA, IBM, Reflection, Liquid and llama.cpp performance evidence remains largely publisher-reported. GitHub license-compliance and Rosalind pages lacked verifiable hour-level timing, so date fields were not forced inside this run.