bit.fanBit.FanWeb3 News · Betting Reviews · Market Insights

SWE-Bench

Jevgrep
2026-09-28 10:43:08

Jevgrep targets code search for coding agents, with SWE-bench run showing a 28.63% cost drop

Developer David has open-sourced Jevgrep, a research tool built on Jev to help coding agents find relevant code before handing the context to tools such as Claude Code and Codex. The setup is aimed at the code-search step, which the project says is one of the most token-intensive parts of agent workflows. Rather than feeding an entire repository into a model or relying on a single semantic search pass, Jevgrep searches a codebase in stages, decides which directories are worth exploring, inspects related files and declarations, and returns source snippets, line numbers, and clues for follow-up reading. The repository also includes a Skill that tells an agent when to call the `jg` command to gather context. In the latest SWE-bench test covering 10 Python tasks, runs with and without Jevgrep both solved 8 tasks, but total GPT-5.6 Sol cost fell from $7.62 to $5.44, a 28.63% reduction. The author had initially posted 40% on X, but the repository later updated the result to about 30%. That figure only covers Sol costs. Logged Jev calls account for at least another $1.57, while some calls do not have complete billing records, leaving the actual total unknown. The test set was also limited to 10 tasks with one run per task.

450
Jevgrep targets code search for coding agents, with SWE-bench run showing a 28.63% cost drop
artificial in
2026-09-03 04:18:25

Stanford HAI 2026 AI Index Report: AI Transitions from Information Generation to Task Execution, US-China Gap Narrows to 2.7%

The Stanford HAI released the Chinese version of its 2026 AI Index Report, highlighting a shift from model capability competition to infrastructure and institutional transformation. Leading models approach 100% human performance on SWE-bench, while agent task success rates jumped from 12% to 66%. The gap between top US and Chinese models narrowed to 2.7%, with frequent leadership swaps. Global AI data center power capacity hit 29.6 GW, and chips, energy, and water are becoming critical bottlenecks.

1000
Stanford HAI 2026 AI Index Report: AI Transitions from Information Generation to Task Execution, US-China Gap Narrows to 2.7%
IBM
2026-08-26 05:50:50

IBM unveils Granite 4.2 open-source reasoning models with native chain-of-thought support

IBM has released the Granite 4.2 family of open-source reasoning language models, introducing three parameter sizes: 3B, 8B, and 30B. The series is built around explicit reasoning, with each model able to generate a chain of thought before answering. IBM also included a thinking/non-thinking switch and a low-power mode across the lineup. The models use a decoder-only dense Transformer architecture, were pretrained from scratch on roughly 15 trillion tokens, and then post-trained through a multi-stage reinforcement learning pipeline. IBM said the 8B and 30B versions include agentic reinforcement learning modules, allowing them to learn tasks such as editing code, operating a terminal, and performing web search in real sandbox environments. All Granite 4.2 models are released under the Apache 2.0 license, which permits downloading, fine-tuning, and commercial production use. IBM also introduced two 470M-parameter Granite Speech 5.0 Turbo CTC speech models. In SWE-Bench Verified, the 8B and 30B models scored 47.67 and 57.00, respectively. According to MarkTechPost, the models are aimed at sectors including software development, finance, healthcare, telecommunications, and the public sector.

1130
IBM unveils Granite 4.2 open-source reasoning models with native chain-of-thought support