AI Daily May 27, 2026: Models Compete on Benchmarks, Agents Compete on Scaffolding

Forum topic · 小凯 · 2026-05-31

Summary

A Chinese tech forum's AI daily roundup for May 27, 2026 covering: Qwen 3.7 Max's coding-focused release (4th on Code Arena Frontend, near Claude Opus 4.6) with community caveats that closed API performance may not match open weights; Gemini 3.5 Flash hitting ~280 tok/s but with ~5x cost increase and flawed default thinking modes; Claude Mythos solving Erdős Problem 90, reinforcing that external harnesses unlock research-level math performance; the emerging consensus that agent competition centers on execution frameworks, verification loops, and memory (DeepSWE, Claude Code security plugins, /workflows, W&B MCP Server); vLLM's new Rust frontend boosting preprocessing-heavy throughput from ~162 to ~837 req/s; 800VDC data center power trends; the 'Language Models Need Sleep' context compression paper; QUEST open deep-research models; OpenRouter's $113M Series B; and DeepSeek's reported ~$10.29B funding round.

Today's AI news is dense. Here are the threads most worth following.

🔥 Model layer: Qwen 3.7 Max and Gemini 3.5 Flash

Qwen 3.7 Max is out, officially focused on coding, software engineering, tool calling, and reasoning. It ranks 4th on Code Arena Frontend, close to Claude Opus 4.6. But the community warning is spot-on: the performance of a closed-source API does not guarantee the later open-weight release will be equally strong. This matters — too many people have been misled by API-only scores.

Gemini 3.5 Flash is fast: Artificial Analysis measured roughly 280 tok/s, with stronger agent performance. The cost, though, is about 5x higher. The subtler issue is product-level: the Gemini App in default mode even fumbles 300+140, and only works properly with higher thinking settings. This reflects a trend: a model's "default personality" is increasingly a product problem, not just a technical one.

🧠 Math spillover: Claude Mythos solved Erdős Problem 90

Several researchers discussed Claude Mythos's math performance. Bubeck's core point: with a well-built external harness, Mythos and GPT-5.5 can deliver research-grade results far beyond what the ordinary chat interface shows.

The implication: a model's capability ceiling depends heavily on how you scaffold it. A bare chat interface and a system with tool calling, verification loops, and iteration space may invoke the same model — but the output can differ night and day.

🤖 Agents and tooling: the harness is the new battleground

The coding-agent consensus is converging: the competition is no longer just about base models, but about outer execution frameworks, verification loops, memory, and routing. Discussions from DeepSeek, Google Managed Agents, LangChain, and dair.ai all point this way.

Notable moves:

DeepSWE — called one of the few benchmarks that genuinely approximates real coding experience, and good at separating top models
Claude Code security plugins — after Anthropic internal use, security-related PR review comments dropped 30%–40%
Claude Code /workflows (tested then pulled) — shifts multi-agent orchestration from "let the model improvise" to code-style编排-free declarative orchestration, with stages, concurrency, retries, and budget control
W&B MCP Server — lets coding agents read experiment records directly, with a schema-first design that avoids blowing up the context window

One takeaway: agent competition in 2026 has moved from "whose model is smarter" to "whose scaffolding is sturdier".

⚡ Infrastructure: Rust frontends and high-voltage power

vLLM added a Rust frontend replacing the Python API server. In preprocessing-heavy scenarios, request throughput rose from ~162 req/s to ~837 req/s. This is especially useful for deployments bottlenecked at the CPU/API layer.

Data centers are looking at 800VDC. SemiAnalysis discusses borrowing high-voltage power-electronics approaches from EVs to support denser AI server rooms. As GPU power keeps climbing, power delivery and distribution are no longer edge concerns — even John Carmack shared this one.

Epoch AI's warning: inference compute demand may already be outpacing existing supply, especially for long-context tasks.

🔬 Research: a new angle on context compression

The paper "Language Models Need Sleep" is interesting. It likens long-context handling to memory consolidation during sleep: first transfer recent content into persistent fast weights, then clear the KV cache. Long tasks then don't have to carry an ever-growing context — very useful for agents running long pipelines.

QUEST open-sourced 2B–35B deep-research models focused on long-horizon factual retrieval, citation alignment, and report generation. The CUSP evaluation found that models can propose research directions, but remain weak at judging "when a breakthrough happens".

💰 Industry: OpenRouter raises $113M Series B

OpenRouter closed a $113M Series B; weekly tokens processed jumped from 5 trillion to 25 trillion in six months. The "multi-model routing layer" is becoming a standalone business.

DeepSeek is reportedly raising ~$10.29 billion, with Liang Wenfeng signaling continued commitment to AGI and open models. A strong signal from the open-source camp.

---

Today's theme in one sentence: models compete on ceilings, agents compete on scaffolding, infrastructure competes on efficiency, and money is competing over the routing layer. Everyone is looking for their own moat.

*(Original post tags: easy-learn-ai, daily updates, memory)*