TL;DR: A quiet day that wasn’t—open-weight models surge, interpretability gets infra, and evals expose agent brittleness

Major Highlights:

• Open-weight models tighten the gap (GLM‑4.7, MiniMax M2.1): GLM‑4.7 landed with day‑0 ecosystem support across MLX, vLLM, Ollama, and agent stacks (TRAE). Early reports show solid local throughput (≈16 tok/s on MLX with batching gains) and strong coding performance; ValsAI ranks it #1 open-weight with a reported +9.5% over GLM‑4.6, and Deedy cites 73.8% on SWE‑Bench. MiniMax’s M2.1 (230B total / 10B active MoE) positions as a Claude‑like open alternative for coding/agents with 200K context and large max outputs, emphasizing workflow fit (orchestration, “deep research agents”) over leaderboard chasing.
• Interpretability becomes infrastructure (Gemma Scope 2): Google DeepMind released SAEs and transcoders for every layer across all Gemma 3 models (270M–27B, base+chat), lowering the marginal cost of mech‑interp analyses. Community leaders frame this as a reusable substrate for safety, debugging, and feature-level probing (Neuronpedia highlighted).
• Benchmarks spotlight reality checks (medicine, ARC, APIs): Medmarks v0.1 debuts as an open medical eval suite/leaderboard spanning 15+ environments with verifier-based scoring. ARC‑AGI discourse intensifies, with Poetiq reporting up to 75% on ARC‑AGI‑2 using a custom harness with “GPT‑5.2 X‑High” at <$8/problem—prompting concerns about harness-overfitting and generalization. WAPIIBench shows LLMs struggle with real API invocation (Asana, GCal, Sheets, Slack): open models solve <40% of tasks; regex constraints from OpenAPI specs eliminate illegal calls and markedly raise correctness.

Key Technical Details:

• GLM‑4.7: MLX interactive speeds ≈16 tok/s; vLLM day‑0 support includes MTP decode, tool/function calling, and “thinking controls”; available via Ollama and agent stacks; ValsAI +9.5% vs GLM‑4.6; SWE‑Bench 73.8% (reported).
• MiniMax M2.1: 230B total parameters with 10B active (MoE); 200K context; strong SWE-* and internal “VIBE‑bench” claims; early adoption via Cline/Ollama-like tooling.
• Gemma Scope 2: SAEs + transcoders trained for every layer across Gemma 3 sizes (270M–27B), base and chat variants.
• Medmarks v0.1: Open medical evaluation suite/leaderboard; 15+ environments; verifier-based scoring.
• WAPIIBench: <40% solve rate on 4 real APIs; regex constraints from OpenAPI specs reduce illegal methods/URLs/args to zero and boost task correctness.
• Agents/workflows: Vercel simplified its text‑to‑SQL agent by removing ~80% of tools and adding a sandbox—40% fewer tokens, 40% fewer steps, 3.5× faster execution.
• Other: Qwen touts “R3” (Rollout Routing Replay) in SGLang; ongoing chatter on long‑RL/GRPO and post‑training divergence.*

Community Response/Impact:

• “Best open source model” narratives coalesce around GLM‑4.7; personality framing emerges: “GLM ≈ open GPT, MiniMax ≈ open Claude.”
• Interpretability release seen as a turning point—shared probes enable reproducible, safety-relevant analyses without bespoke training.
• Skepticism on ARC gains via harnessing and prompting; renewed interest in constraints and structured decoding for agents.

First Principles Analysis:

• Open-weight momentum is being driven by ecosystem readiness (MLX/vLLM/Ollama) and MoE economics (10B active) that deliver capability at lower runtime cost.
• SAEs/transcoder artifacts shift interpretability from bespoke research to reusable infrastructure, enabling systematic auditing and feature-level controls.
• Agent reliability improves more via simplification (fewer tools, sandboxing, constrained decoding) and better state representations (call‑stack contexts) than via tool sprawl.