Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-30 · high
Frontier models solve fewer than 1 percent of these tasks, and MindForge's gains carried over to seven other benchmarks too.
Why now: The paper posted to arXiv in July.
Qwen3.6-27B jumped from solving 37.98% to 49.51% of tasks on a from-scratch coding benchmark, per a MindForge paper posted to arXiv in July.
That matters for anyone building coding agents. Frontier models resolve under 1% of these tasks, so a 27B open model closing that gap changes which teams can compete without frontier-scale budgets.
The task is designed to resist shortcuts. Agents write code from scratch against only natural-language documentation, checking their work against an execute-only binary with no source to read or copy.
MindForge's fix isn't a bigger model. Its system auto-converts open-source command-line programs into training environments. GLM-5.2 then serves as a teacher, generating synthesis trajectories the smaller model learns from, per the paper.
The gains didn't stay confined to that one benchmark. Tested against seven unseen benchmarks, per the paper, results included gains of 31.00 points on RepoZero-C2Rust and 14.16 on DeepSWE. NL2Repo-Bench rose 10.70, SWE-bench Pro climbed 5.93, and SWE-bench Verified gained 5.04 points.
Each link below shares sources, entities, or timing with this story.
The 397B MoE scores 86.1 on Terminal-Bench 2.1, 86 on SWE-Bench Verified, 56 on DeepSWE, 92.8 on GPQA Diamond and 44.6 on HLE, which the team frames as comparable to Claude Opus 4.8. The method is a closed self-improvement loop where the model proposes its own tasks and scaffo...
The 397B MoE scores 86.1 on Terminal-Bench 2.1 (Terminus-2) against Claude Opus 4.8's 85.0, 86.0 on SWE-bench Verified, and 92.8 on GPQA Diamond. Hugging Face But it trails badly on the harder agentic rows: 13.5 versus 21.1 on Frontier-Bench v0.1, 59.5 versus 69.7 on NL2Repo....
Agent Lightning v1.0 (arXiv 2608.17528) inverts the standard agentic RL architecture, and the inversion is the whole point. Normally the training engine owns the environment loop. It drives the agent, collects trajectories, computes rewards. Which means your training setup and...
SWE-Touch mines task-critical regions from repair trajectories, builds plausible Counter-Edits that conflict with task completion, and injects them with contextual user messages when the agent reaches that code. Across nine models on SWE-bench Verified, average resolve rate dr...
Unisound's U2 is a 266B-total / 10B-active MoE tuned for agents, citing 72.2% SWE-bench Verified at $0.15/$0.30 per 1M tokens, while GLM-5.2 is getting named the strongest open-weight coding model across July roundups. Both are roundup-sourced, so verify the benchmarks against...
Everyone benchmarks per task. Accuracy on SWE-bench, pass rate on Terminal-Bench, a leaderboard row per model. Together AI ran the experiment sideways: fix the budget at $100, point both models at DeepSWE, and count how much work came out the other end. GLM-5.3 finished 17 tas...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.