Daily Executive Summary

  • DeepSeek opens low-cost Search and Fetch APIs: each costs $0.75 per 1,000 successful non-empty calls; 5 QPS limits and vendor relevance labels constrain use as a general truth source.
  • Dory1.0-4B offers adjustable compute depth from one checkpoint: one, two or four loops trade accuracy for latency, while stable vLLM and llama.cpp support is not ready.
  • DeepSeek V4.1-Flash-1017 updates both weights and the hosted alias: interfaces and pricing are unchanged; vendor-reported coding gains need same-task A/B reproduction.
  • RTX Spark brings large-memory CUDA development to Windows PCs: top configurations advertise 128 GB unified memory, but SKU-level stock, pricing and independent performance are inconsistent.
  • FDE-Gym moves agent evaluation from issue fixing to customer delivery: 95 tasks require discovering unstated constraints; the best configuration resolved only 14, while the public harness includes three full examples.

Models

  • Dory1.0-4B (new / TRIAL): its shared 20-layer recurrent block runs 1/2/4 times. Recurrent depth means reusing the same layers on each token; the hard part is that higher accuracy also costs more compute and current runtimes need unmerged code. Model card
  • DeepSeek V4.1-Flash-1017 (updated / TRIAL): MIT weights and the deepseek-flash hosted alias are updated with unchanged price and interface. Terminal-Bench and other results are vendor-run under different harnesses. Release

Agent & AI Engineering

  • DeepSeek Search / Fetch (new / TRIAL): Search returns structured results and Fetch turns one page into Markdown. The authority field is retrieval metadata, not a certification of factual trust. Guide
  • Claude Code 2.1.302 (updated / ADOPT): fixes full routine replay after partial failure, lost config writes and MCP compatibility. Routines with external side effects need failure-injection regression. Stable release
  • Open WebUI 0.12.3 (updated / TRIAL): adds temporary MLS end-to-end encrypted chat, lighter Redis streaming and workspace search. MLS coordinates group-session keys; deployment still requires HTTPS, WebSockets and disabled server-side WebSocket logging. Stable release
  • GPT-Live prompt-to-voice (updated / TRIAL): all API customers can create voices from text descriptions. Creation is free and speech uses GPT-Live billing; the voices work only with GPT-Live and creation is incompatible with ZDR. Documentation

Open Source

Dory provides four weight formats, FDE-Gym provides an Apache-2.0 harness and three examples, and DeepSeek 1017 provides an MIT checkpoint. Downloadable artifacts do not reproduce every claim: stable runtimes, full benchmark data and independent evaluation remain the main gaps.

Research

  • FDE-Gym (new / TRIAL): 95 tasks cover 20 scenarios and 18 sectors, scored by customer business uplift minus contact cost. An unstated constraint is a rule found only by asking the simulated customer; this is the part code benchmarks usually omit and real delivery often fails on. Paper · Code

Developer Tools

  • GitHub License Compliance (new / TRIAL): GA for Enterprise Cloud customers with GitHub Code Security, with rulesets that can block merges by license. The official page is date-only, so this is a late discovery from first observation after the prior run, not an hour-level publication claim. Announcement

Infrastructure

  • RTX Spark (new / WATCH): NVIDIA and OEMs launched Arm Grace + Blackwell RTX Windows 11/WSL PCs. Store links coexist with availability alerts, so exact SKUs need checking. Launch

Business & Policy

No new major policy change entered the event library. No new document materially changed the prior judgment on China's AI technology-export draft and temporary registration.

Trend Signals

  • Trend #1: DeepSeek 1017 adds a public checkpoint but no independent coding benchmark; strengthening / Medium is unchanged.
  • Trend #3: Claude Code default fork subagents and Open WebUI collaboration controls further support convergence on multi-agent runtimes; established is unchanged.
  • Trend #5: Dory and RTX Spark add local-execution paths, but software and supply remain early; candidate is unchanged.
  • Trend #7: routine replay fixes and encrypted collaboration further support separating execution capability from consequence control; strengthening is unchanged.
  • No sufficiently evidenced new trend emerged in this cycle.

Tech Radar

  • New · DeepSeek Search (ai-engineering / TRIAL): low-cost and success-billed; validate relevance, attribution and injection risk.
  • New · Dory1.0-4B (foundation-model / TRIAL): selectable depth is testable, while stable runtime support is not ready.
  • DeepSeek V4.1 Flash (foundation-model / TRIAL): 1017 updates weights and hosting; reproduce vendor gains on the same tasks.
  • New · RTX Spark (infrastructure / WATCH): launched, but buying needs SKU-level stock and independent performance.
  • New · FDE-Gym (research / TRIAL): public examples enable a delivery-evaluation trial; the full data is controlled.
  • New · License Compliance (developer-tools / TRIAL): merge blocking is GA; test plan limits, false positives and exceptions.
  • Open WebUI 0.12.3 (developer-tools / TRIAL): encrypted collaboration and Redis improvements are testable; regress Skill prompt cost.
  • Claude Code 2.1.302 (developer-tools / ADOPT): replay and recovery fixes support upgrading.
  • GPT-Live (ai-engineering / TRIAL): text-described voices are GA with product and retention boundaries.

Current radar: ADOPT 9 / TRIAL 127 / WATCH 56 / IGNORE 7. See the site Insights page for the full current radar.

Worth Trying

  • Compare DeepSeek Search with the current retriever on the same queries, recording usable-result rate, Fetch success and cost per success.
  • Run the same coding tasks with Dory at 1/2/4 loops, recording success, wall time and peak memory.
  • Use FDE-Gym's three public cases to measure business uplift, incidents and contact cost for the current coding agent.

Sources

Coverage note: all 36 fixed monitors were fetched and reviewed; active discovery added Chinese primary material, small teams, open source, research, infrastructure, policy and exceptional general technology. Some DeepSeek and OpenAI pages are date-only, and GitHub license compliance lacks an hour-level timestamp; vendor benchmarks and RTX Spark stock still need independent or SKU-level validation. No general-news candidate met the exception threshold.