"Bearly OpenADE GPT-5.5, Alibaba Qwen3.7 Max, Datacurve DeepSWE" — AlphaSignal

Why this is in the vault

Two front-page items cross the RDCO threshold hard. Qwen3.7 Max ships a 1M-token context window plus Anthropic-API-protocol compatibility (drop-in to Claude Code), which is both a context-management and a substrate-portability data point. And a research item teaches models a "sleep" method to compress long context without slowdown — direct convergence with the "Language Models Need Sleep" item already flagged this morning and with RDCO's own context-rot discipline. Lior's editorial framing this issue is endurance over raw intelligence: "Longer context, longer shifts, longer tasks." That is exactly the axis RDCO's always-on COO agent lives on.

⚠️ Sponsorship

This issue carries three paid placements plus the standing house ad, all disclosed here:

Viktor — "Presented by" placement under the OpenADE top-repo item. Pitch: company-wide AI-native agent across 3,000+ tools from Slack/Teams; $100 credits CTA. Third-party advertiser.
Lambda — "Presented by" placement under the Qwen top-news item. Pitch: GPU training infrastructure ("most AI teams treat compute as a commodity"). Third-party advertiser.
ASUS — "Presented by" placement, also runs as Signal #2 ("ASUS NUCs ... AI mini PCs"). Third-party advertiser.
"Work With Us" — AlphaSignal's standing house slot soliciting advertisers to reach 250,000+ AI developers.

None of the sponsors overlap with the editorially-covered items below. Treat sponsor copy as ad, not signal.

Issue contents

Curation issue (read time ~6.5 min), curated by Lior Alexander, founder of AlphaSignal and former ML engineer.

Top Repo — Bearly AI ships OpenADE (63K likes). Open-source local agentic coding loop: describe task, get file-level plan with edge cases, comment/refine/lock, one-click execute with automatic git snapshots for rollback. Now runs on GPT-5.5/Codex (claimed 82.7% on Terminal-Bench 2.0, fewer tokens) and supports Claude Code. Free, local, code never leaves the machine.

Top News — Alibaba Qwen3.7 Max (3.6K likes). Reasoning model into the "Go" tier. 1M-token context window (~2,000 pages, fits an entire codebase). Text-only I/O. Claimed 80.4% SWE-Bench Verified, 92.4% GPQA Diamond. Runs autonomously up to 35 hours / 1,000+ tool calls. Supports the Anthropic API protocol, so it drops into a Claude Code setup by swapping the model. Proprietary, API-only, no open weights.

Top Repo — Datacurve DeepSWE (2.5K likes). New coding benchmark built from scratch to expose model spread that existing leaderboards hide (public-GitHub contamination, tasks too small, noisy grading). 113 tasks across 91 repos in TypeScript/Go/Python/JS/Rust. Claimed result: GPT-5.5 leads at 70%, 16 points ahead of next model.

Signals: (1) Figure robots heading to JCPenney and Brooks Brothers stores, sorting 88K packages. (2) ASUS NUCs [sponsor]. (3) New "sleep" trick lets language models compress long context without slowing down. (4) Nango open-sources the API integration layer "SaaS companies pay $50K a year to rent." (5) Free open-source tool forces Claude Code and Cursor to code "like a senior dev." (6) DeepSeek infrastructure papers called the biggest open-source AI contribution this century.

Mapping against Ray Data Co

1M-token context (Qwen3.7 Max) → context-rot discipline, not a free win. The headline "drop an entire codebase in" is exactly the temptation CLAUDE.md rule #4 warns against. More context is not free — model performance degrades as context grows. A 1M window does not retire the subagent-routing rule; it raises the ceiling but the same context-rot tax applies. RDCO's posture stays: route long artifacts through subagents, don't dump them into parent context just because the window can hold them. The Qwen number is evidence the industry is selling the window-size frame; RDCO's discipline is the counter-frame.

Anthropic-API-protocol drop-in → substrate-portability brief, confirmed. Qwen3.7 Max plugs into Claude Code by swapping the model, no agent rewrite. This is a concrete data point for the substrate-portability thesis (see [[2026-04-24-gpt-5-5-workspace-agents-substrate-threat]]): the harness layer is becoming model-agnostic, which cuts both ways for RDCO — lower switching cost if Anthropic stumbles, but also commoditizes the substrate the COO agent runs on. Worth tracking whether the protocol-compat trend strengthens RDCO's portability optionality or erodes its moat.

"Sleep" memory-compression method → triple convergence. This item converges with (a) the "Language Models Need Sleep" item in this morning's frontier digest, and (b) the context-rot discipline in rule #4. The shared thesis: long-running agents need an active memory-management mechanism, not just a bigger buffer. RDCO already implements a manual version of this — subagent extraction, working-context.md as durable scratchpad, compaction summary rules. If model-level memory compression matures, it could reduce RDCO's reliance on manual subagent routing, but until then the discipline holds. This is the on-thesis item to watch.

DeepSWE benchmark spread → model-selection signal. A benchmark that actually separates models (16-point GPT-5.5 lead) is more useful than saturated leaderboards for RDCO's model-routing decisions (which model for which agent tier). File as a candidate input, not yet acted on — claimed numbers are vendor/curator-reported, unverified.

Mapping strength: strong on the Qwen 1M-context + sleep-compression items (both tie directly to load-bearing RDCO discipline and an active brief). Endurance-over-intelligence framing is the throughline.

Curation section — notes

All performance numbers (82.7% Terminal-Bench, 80.4% SWE-Bench, 92.4% GPQA, 70% DeepSWE, 35-hour autonomy) are vendor- or curator-reported and unverified — treat as claims.
Deep-fetch cap (2 third-party links) not exercised: the two on-threshold items (Qwen context, sleep compression) are summarized sufficiently in-newsletter and the sleep paper is already captured via the Innermost Loop digest. No external fetch added value worth the context cost.
The "sleep" Signal is the same underlying research as the frontier-digest "Language Models Need Sleep" item — same convergence, cross-linked below rather than re-fetched.

[[2026-05-27-innermost-loop-may-26-frontier-digest]] — same-morning digest; "Language Models Need Sleep" item is the same memory-compression research surfaced here as a Signal
[[2026-04-24-gpt-5-5-workspace-agents-substrate-threat]] — substrate-portability brief; Qwen's Anthropic-API-protocol drop-in is a fresh data point
[[2026-05-26-skillopt-self-evolving-agent-skills]] — adjacent agent-capability thread (self-evolving skills) on the endurance/autonomy axis
[[2026-04-20-alphasignal-claude-design-word-lingbot-map]] — prior AlphaSignal issue under the same watch