Fetching from the wire…
01
02
03
04
05
06
07
08
Models2026-09-16 · source-backed
The Apache-2.0 repo was last modified September 16 with 1,048 downloads against paper arXiv:2609.11977. It's a Qwen3.6-35B-A3B post-train, and the pitch is cost-per-episode rather than peak capability: co-work agents spend most steps on state tracking, coordination, recovery and follow-through, not frontier reasoning, so they trained execution-grounded replayable long-horizon trajectories across multiple harnesses. Relevant if you're paying for a long agent loop rather than a single hard question.
Each link below shares sources, entities, or timing with this story.
Accio open-sourced it August 2, now 1,042 stars: 107 tasks across 14 containerized mock services replicating Alibaba, Shopify, and FreightOS (GitHub). Task mix is 53 CLI, 28 browser, 16 file ops, 10 API/MCP, split 65 text-only / 20 browser-text / 22 vision-requiring. Twelve mo...
Not mocks, not static transcripts. High-fidelity reproducible replicas of real online commerce services, with a reproducibility contract, live leaderboard, and reference results run through the OpenClaw harness. The stateful-replica approach targets the failure mode static ben...
The leaderboard says first place. The methodology says you should check your own bill. Qwen3.8 Max now ranks first on Artificial Analysis' agentic index, scoring 86.1 on OSWorld-Verified ahead of GPT-5.6 Sol Max at 83.2 and Fable 5 at 85.0, priced at $2.00/M input and $6.00/M...
Edge0-AI/Edge0, created September 8, went from 269 to 583 stars in two days. It packages SSD expert offload, Recover-LoRA and prerouter routing prediction into an MLX-backed framework: edge0-35b is a 4-bit 40-layer 256-expert model built on Qwen3.5-MoE 35B-A3B needing about 2....
Created August 24, it holds a trendingScore of 3,967 against second-place GLM-5.3-Flash at 1,376 (Hugging Face). The near-1:1 like-to-download ratio means almost everyone bookmarking it hasn't pulled weights, and the unsloth GGUF conversion at 4,354 downloads is absorbing comp...
A Hugging Face post dated September 10 documents 96+ hours across 1,000+ quantization configurations on Qwen 3.5 0.8B and 4B, producing per-tensor layout maps replacing the generic GGUF heuristics. Findings: token embeddings are 8-16x more sensitive to degradation than other w...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.