AI Trend Notifier
EN한
← archive

$ cat briefs/daily/2026-10-18.md

2026-10-18

October 18, 2026 (Sun)

2 stories · 2 paper picks · 4 new pages

+4new pages
[01]

Top Stories

1. Beam moves from preview to downloadable weights and agent tooling

Reflection released its text-only model under Apache 2.0, with 501.3B total / 23.4B active parameters, 1M-token context, and BF16, FP8 and NVFP4 weights. Mirror CLI adds tool approvals and resumable work; self-distillation recipes are still promised.

Why it matters: developers get deployment and customization paths alongside the checkpoint; selected demos do not establish production reliability. → Beam 1.0 (release) (model card)

2. DeepSeek updates the model behind deepseek-flash

The 2026-10-17 update reports Terminal-Bench 4.0 42.3, DeepSWE v1.1 77.0 and ProgramBench 29.1. The first two use Claude Code and Mini-SWE-V2, respectively. API calling method and pricing are unchanged.

Why it matters: a stable API alias now serves a different checkpoint; results need checkpoint and harness identification. → DeepSeek V4.1-Flash (source)

[02]

Paper Picks

OpenGCUA: combining code execution with visual interaction raises mean OSWorld-v2 partial score by +19.6 points across eight models, with 53.0% fewer reported steps than batched GUI-only execution. Read it for interface choice; partial credit is not completion. → OpenGCUA (source)

FDE-Gym: the best realistic-setting configuration resolves only 14 of 95 simulated delivery tasks; its 0.56 mean score measures penalized, normalized customer-metric improvement. Read it for discovering unstated constraints and then executing successfully. → FDE-Gym (source)

[03]

Watch

  • DeepSeek Search and Fetch cost $0.75 per 1,000 successful requests each. Failed requests and empty searches are unbilled. Server-side coding-harness calls start billing at 2026-11-01 00:00 Beijing time. → Agents (LLM Agents) (source)
[04]

New in Wiki

  • Reflection — new organization page for reader review. (source)
  • Beam 1.0 and the two paper notes above add specifications, methods and evidence limits. (source)
[05]

Updates