Daily Executive Summary

  • Named GPT-4 API snapshots now have a retirement date: OpenAI will remove listed GPT-4o, GPT-4o mini and GPT-4.1 snapshots, including fine-tuned variants, on January 14, 2027 and recommends gpt-6-luna. The old snapshots remain callable; teams need an exact model-ID inventory rather than treating this as an immediate shutdown.
  • NVIDIA opens the full Run:ai platform: The Apache-2.0 repository covers the control plane, scheduling, GPU fractioning, CLI and MCP server. Shared-GPU operators can inspect the whole stack, but the community build has no enterprise support and the NGC distribution carries separate NVIDIA terms.
  • LFM2.5-24B-A2B adds a local-agent option: Liquid AI released a 24B-total MoE that activates about 2B parameters per token with broad day-one runtime support. Performance is vendor-reported, and entities earning at least $10 million annually need a separate license.
  • Claude Code 2.1.299 closes execution boundaries: The stable release fixes sandbox writes into plugin directories, duplicate resumed execution, MCP startup before policy checks and third-party gateway-key leakage. Teams using managed policy or custom gateways should prioritize the upgrade and replay those paths.
  • Third Moment reports a 10,000-agent long-running system: An OpenAI customer story says the firm runs up to 10,000 agents for 36 to 48 hours with outer models supervising stalled work. It is a rare architecture signal, but raw telemetry and independent validation are absent.

Models

  • GPT-4 API snapshot retirement (new / ADOPT): Affected snapshots remain callable; shutdown is January 14, 2027. The page gives a notification date but no reliable announcement hour, so migration must follow the exact listed model IDs and fine-tuned assets. Official documentation
  • LFM2.5-24B-A2B (new / TRIAL): The 24B MoE activates about 2B parameters per token and ships base and post-trained weights. MoE calls only part of its expert network for each token; the hard part is that sparse activation does not eliminate memory, routing or license costs. Release post · Model card
  • Gemini 4 Argon (updated / TRIAL): The model moves from controlled preview to a paid API. The unified token budget means input and output share the one-million limit. Long runs can increase output, continuation, recovery and human-review cost together. Official post
  • Grok 4.7 (new / TRIAL, late discovery): The public API supports text, images, tools and 500K context. Input and output rates double above 200K prompt tokens, and the US regional endpoint adds 10%. Announcement · Documentation
  • MiniCPM-V 4.7 (new / TRIAL, late discovery): The 35B-A3B and 1B weights use Apache-2.0. Expert lookahead preloads likely next-layer experts without changing actual routing. It reduces memory stalls but not storage, bandwidth, heat or battery cost. Model card
  • GLM-Legal (new / WATCH): The GLM-5.3-Flash-based legal model is in an application-only beta. Harvey LAB all-pass is 19.2%, with some evaluation run by Vals. There are no public weights or generally available API, and AutoClaw support is still forthcoming. Official post

Agent & AI Engineering

  • Third Moment Agent System (new / WATCH): The customer reports up to 10,000 concurrent agents running for 36 to 48 hours. Outer supervisors terminate, retry or escalate stalled workers; concurrency is not effective throughput, and success, cost and recovery telemetry are absent. Customer story
  • Claude Code 2.1.299 (updated / ADOPT): Adds idle compaction and runner drain while fixing sandbox, resumed-execution, MCP-policy and gateway-credential boundaries. Stable release
  • Gemini CLI 0.64.0 (updated / WATCH): Atomic file operations, state backup recovery, Ctrl+C cancellation, ACP usage updates and A2A settings migration reach stable; A2A still lacks a formal compatibility commitment. Stable release
  • ColGrep-Agent (new / TRIAL): A local search model locates relevant code before passing smaller context to a coding agent. File F1 measures whether it found the right files; it does not show whether the final patch passes tests, so evaluation must remain end to end. Release post · Code

Open Source

The complete Run:ai GPU orchestration stack is now open source; LFM2.5, DoubtLess and Hybrid Agent also ship downloadable weights or code. GitHub secret scanning now supports Anthropic service and user API keys in GA and blocks them by default with push protection. This is prevention, not evidence of a leak. GitHub announcement

ColGrep-Agent and StudentSim both provide runnable artifacts. No new stable general-purpose agent runtime cleared the bar; window candidates from Gemini CLI, Qwen Code, Codex and Ollama were nightlies, alphas or a release candidate.

The recall audit also records four high-attention but under-evidenced items as IGNORE: Janas has author-only performance comparisons; JEV-35B self-scores a public subset outside the official board; Agent37 has only a launch discussion; and the EmbeddingGemma 2 transcription demo provides no public training code or test set. Attention is recorded without turning it into a trial recommendation.

Research

  • DoubtLess (new / TRIAL): A small model reads the same chain of thought under brief and verbose instructions, and their logit difference steers the larger model. The authors report lower reasoning-token use, but there is no independent reproduction and the reference setup needs an 80 GB GPU. Paper · Code
  • Hybrid Agent (new / TRIAL): HybridLink switches between hidden latent reasoning and explicit tool interaction while carrying state across the boundary. Public 8B and 14B weights make it testable, but the 11-benchmark results are author-reported. Paper · Code
  • StudentSim (new / TRIAL): The paper covers 60 learners across chess, second-language learning and mathematics. Held-out evaluation tests interactions the simulator did not see during training. Better fit on unseen records still does not mean real students learn better. Research post · Code · Paper
  • Encrypted capability backdoors (new / WATCH, late discovery): The paper constructs digital lockers that encrypt hidden computation in weights and unlock it only under a secret trigger. It proves a worst-case limit for elicitation audits, but experiments use a small chemical-reaction proxy. Paper

Developer Tools

GitHub's default Anthropic-key blocking is a low-cost ADOPT item. Claude Code and Gemini CLI stable releases respectively address execution boundaries and state integrity; upgrade testing should include real interruption, recovery and policy-failure cases.

ColGrep-Agent is the run's most actionable developer-tool signal. It searches repositories locally with read-only access and documents Codex, Claude Code and OpenCode integrations. The name says 2B while the model card reports roughly 3B parameters, so deployment should verify actual package and memory requirements.

Infrastructure

  • NVIDIA Run:ai (new / TRIAL, late discovery): The Apache-2.0 repository contains the control plane, cluster services, installer, CLI, MCP server and four existing components. Before replacing a supported platform, test tenant quotas, fractional-GPU isolation and upgrades in an isolated cluster. Announcement · Code
  • Maia 300 (new / WATCH): Microsoft plans production deployments from 2027. A 96-accelerator rack has 27.6 TB HBM and 1.7 PB/s aggregate bandwidth. The 40% tokens-per-megawatt and up-to-seven-times rack-throughput figures are Microsoft claims. Announcement · Architecture

Business & Policy

The seven-day review of QGIS's AI-assisted-code ban found no public enforcement case or adoption by another major open-source project, so it remains WATCH. No general-interest event cleared the exceptional-impact threshold this cycle.

Trend Signals

  • Trend #3 gains Third Moment and Hybrid Agent signals: long-running systems are making outer supervision and reasoning-mode switches explicit, but neither signal changes the established status alone.
  • Trend #5 gains the full Run:ai open-source platform. It broadens the self-hosted AI-infrastructure boundary without independent migration or adoption data.
  • Trend #7 gains Claude Code 2.1.299 policy and execution-boundary fixes; agent governance continues moving from prompt conventions to enforceable controls.
  • The other four trends received no evidence sufficient for a status change.
  • No evidence-backed new trend emerged this cycle.

Tech Radar

  • New · GPT-4 snapshot retirement (foundation-model / ADOPT): The shutdown date is fixed; inventory exact model IDs and fine-tunes and validate the replacement on representative evaluations.
  • New · Anthropic key protection (agent-security / ADOPT): Detection and default push protection are GA; keep them enabled and verify organization coverage.
  • Claude Code (developer-tools / ADOPT): Version 2.1.299 closes sandbox, resumed-execution, MCP-policy and gateway-credential boundaries; replay those paths after upgrading.
  • Gemini CLI (developer-tools / WATCH): Version 0.64.0 hardens state and cancellation; A2A still awaits formal documentation and compatibility commitments.
  • New · NVIDIA Run:ai (infrastructure / TRIAL): The full open stack is testable, but the community build lacks enterprise support; start with an isolated-cluster trial.
  • New · LFM2.5-24B-A2B (foundation-model / TRIAL): Public weights and runtime support enable a trial; license limits and vendor benchmarks constrain production judgment.
  • New · DoubtLess (research / TRIAL): Public code supports a one-day rerun; measure task success, tokens and memory together.
  • New · Hybrid Agent (research / TRIAL): Public 8B weights support a small tool-use trial; include failure traceability in the metrics.
  • New · Third Moment Agents (ai-agent / WATCH): The long-running scale is a strong signal, but task-level telemetry and independent validation are missing.
  • Gemini 4 Argon (foundation-model / TRIAL): The public API is available; compare success, drift, review cost and recovery on the same long tasks before and after the price increase.
  • New · Grok 4.7 (foundation-model / TRIAL): Public API and pricing support direct comparison, while vendor benchmark and safety claims still need independent validation.
  • New · ColGrep-Agent (developer-tools / TRIAL): Public weights and integrations support a one-day trial; measure patch success, tokens, latency and total compute together.
  • New · StudentSim (research / TRIAL): The public chess path is reproducible; simulator metrics do not replace trials with learners and educators.
  • New · Maia 300 (infrastructure / WATCH): The roadmap matters for inference planning, but procurement and real-workload evidence must wait for 2027 deployments.
  • New · MiniCPM-V 4.7 (foundation-model / TRIAL): Open weights and deployment code support direct phone tests of quality, sustained speed, memory, thermals and battery use.
  • New · Encrypted capability backdoors (agent-security / WATCH): The worst-case construction changes audit boundaries, but larger-scale independent reproduction is still needed.
  • New · GLM-Legal (foundation-model / WATCH): An application beta and some independent evaluation create a signal; general API access, weights and error reproduction remain absent.
  • New · Janas (open-source / IGNORE): Community attention is high, but independent hardware-matched quality and stability comparisons are absent.
  • New · JEV-35B (foundation-model / IGNORE): Open weights are testable, but the public result is self-scored and not an official board entry.
  • New · Agent37 (developer-tools / IGNORE): Only a founder launch post is available, without sufficient isolation, reliability or adoption evidence.
  • New · EmbeddingGemma 2 STT (open-source / IGNORE): The demo is inventive but lacks reproducible training artifacts and independent word-error comparisons.

Current radar: ADOPT 9 / TRIAL 116 / WATCH 54 / IGNORE 7. See the site's Insights page for the complete current radar.

Worth Trying

  • Inventory production GPT-4o / GPT-4.1 snapshots and fine-tunes, then build a gpt-6-luna comparison and migration date.
  • Deploy the Run:ai community build in an isolated Kubernetes cluster and test quotas, fractional-GPU isolation and rolling upgrades.
  • Trial LFM2.5 on real tool schemas and rerun DoubtLess or Hybrid Agent from public code, recording success, tokens, latency and memory.

Sources

Coverage: all 36 fixed monitors were captured and reviewed, supplemented by active searches across models, developer tools, open source, research, infrastructure, policy and non-English primary sources. The GPT-4 retirement page and LFM2.5 article provide date-level announcement timing only, so this run preserves that limitation rather than claiming hourly precision; Third Moment, LFM2.5 and paper results remain customer-, vendor- or author-reported. Run:ai is an overlap late discovery. Nightly, alpha and release-candidate builds were not treated as stable releases, and no general-interest event cleared the exceptional-impact threshold.