Tuesday, August 4

The signal

1. DeepSeek-V4-Flash-0731 outscores V4-Pro-Preview on every agentic benchmark What happened: DeepSeek shipped the V4-Flash-0731 GA checkpoint on July 31; it scores 82.7 on Terminal-Bench 2.1 vs. V4-Pro-Preview’s 72.1, and 54.4 on DeepSWE vs. 7.3 for Flash Preview. Why it matters: A cheaper flash-tier model beating its own pro sibling on agentic and coding tasks reshuffles the cost/quality tradeoff for anyone running agent workloads on DeepSeek APIs. Confidence: Confirmed Sources: DeepSeek blog (primary) · Artificial Analysis (independent)

2. Inkling-Small: 12B-active open-weights MoE hits 80.2% SWE-bench Verified What happened: Thinking Machines Lab released Inkling-Small on July 30 under Apache 2.0 — 276B total / 12B active parameters, MoE architecture, 1M context, multimodal text/image/audio; it outperforms its larger parent Inkling on coding and reasoning. Why it matters: 80.2% SWE-bench Verified from a 12B-active model under Apache 2.0 is the best publicly available open-weights coding number at this size point, and it runs on commodity hardware via Together AI and Fireworks. Confidence: Confirmed Sources: Thinking Machines (primary) · MarkTechPost (independent)

3. White House finalizes voluntary AI cybersecurity testing framework; Meta at the table What happened: The Trump administration presented a completed voluntary framework for evaluating frontier model cybersecurity capabilities to OpenAI, Anthropic, Google, and Meta on August 3; framework includes a 30-day government early-access evaluation window. Why it matters: Meta’s presence reverses the July TRAINS standoff where they were absent; the framework is opt-in only with no mandatory reporting or public results, but it is the first formal structure for government model evaluations under the current administration. Confidence: Confirmed Sources: Bloomberg (independent) · CNBC (independent)

Since last brief

• model-releases OpenAI cut GPT-5.6 Luna pricing 80% (to $0.20/$1.20 per M tokens) and Terra 20%, and launched Sol Fast mode at 2.5x throughput for 2x cost, effective July 30. VentureBeat (independent)
• agentic-ai EU AI Act GPAI enforcement powers went live August 2; EU AI Office can now request documentation, conduct model evaluations, and fine providers up to 3% of global annual turnover; no formal enforcement actions reported in the first two days. EU AI Act tracker (primary) · Wilson Sonsini (independent)

Models & research

DeepSeek-V4-Flash-0731

Claim: GA release of DeepSeek-V4-Flash; 284B-param CSA+HCA backbone, re-post-trained specifically for agentic tasks. Evidence: Terminal-Bench 2.1: 82.7 (vs. V4-Pro-Preview 72.1); DeepSWE: 54.4 (vs. 7.3 Flash Preview); DSBench-FullStack: 68.7 (vs. 37.0); output speed 113.5 t/s on DeepSeek API. Caveat: All benchmarks are DeepSeek’s own; no independent third-party evaluation published as of this run. Worth your time: Read.

Inkling-Small

Claim: 276B total / 12B active open-weights multimodal MoE (Apache 2.0); beats its parent model Inkling on every coding and reasoning benchmark. Evidence: SWE-bench Verified 80.2% bash-only harness (Inkling: 77.6%); Terminal-Bench 2.1: 64.7%; HLE text 31.6% (Inkling: 29.7%); full weights on Hugging Face, hosted via Together AI and Fireworks. Caveat: Bash-only SWE-bench harness inflates scores compared to standard scaffolding; no independent replication found yet. Worth your time: Read — Apache 2.0 and 12B active parameters make it genuinely self-hostable for coding agent workflows.

GPT-5.6 price restructuring

Claim: OpenAI cut Luna 80% and Terra 20%, and launched Sol Fast mode (2.5x throughput at 2x cost), effective July 30. Evidence: Official OpenAI pricing page; confirmed by independent reporting. Caveat: None noted. Worth your time: Skim if you route agent workloads across GPT-5.6 variants.

Tools worth touching

Panniantong/Agent-Reach

• Does: CLI giving AI agents read/search access to Twitter, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu with no individual API fees
• Why now: 951 stars today; fills the gap for agent-driven research and monitoring without per-platform API keys
• Try if: building agents that need broad web awareness across social and code platforms
• Skip if: scraping third-party platforms creates compliance or ToS risk in your context
• Links: repo

huangruiteng/loopx

• Does: State kernel for long-running AI agent teams; manages durable goals, quota-aware auto-wake, and executable todos across restarts
• Why now: 618 stars today; addresses the persistence gap in multi-agent orchestration where most frameworks assume short-lived runs
• Try if: you’re running agent loops that need to survive crashes or manage budget across extended tasks
• Skip if: your agents are stateless or single-session
• Links: repo

obra/superpowers

• Does: Agentic skills framework and software development methodology for composing agent capabilities
• Why now: 617 stars today; the “methodology” framing alongside the toolkit is an interesting pattern gaining traction in the agent tooling space
• Try if: you want convention-based structure for building and sharing agent capabilities
• Skip if: you need a stable API surface or production support
• Links: repo

Moves on the board

Other

• lyogavin/airllm trending (1,716 stars today): tool for running 70B+ models on a single 4GB GPU; renewed attention likely tied to Inkling-Small’s release and interest in self-hosting large MoEs. repo
• Alishahryar1/free-claude-code trending (449 stars): wrapper claiming free terminal access to Claude Code, Codex, and Pi; signals demand but raises obvious ToS and reliability concerns. repo

Watchlist

• OpenAI ExploitGym / Kill Switch Act: House bill (Lieu-Moran, July 23) still in early committee stage; Senate companion not formally introduced; AISI cross-lab investigation pending (opened 2026-07-23)
• Kimi K3 llama.cpp KDA kernel: PR #26185 still open on ggml-org/llama.cpp; community GGUFs on Hugging Face but require building from the pwilkin branch (opened 2026-07-23)
• Anthropic open-weights position: no new formal statement; attended Aug 3 White House meeting; 77-firm coalition has issued no follow-up (opened 2026-07-30)
• TRAINS: voluntary cybersecurity testing framework finalized Aug 3, all four labs including Meta now engaged; opt-in, no mandatory reporting or public results; watching for framework text publication and lab adoption rates (opened 2026-07-25)
• EU AI Act GPAI enforcement: powers active since Aug 2; no formal enforcement actions in first two days; watching for first investigation or fine notice (opened 2026-07-26)