AI Briefing#open-source
The latest AI and developer news about #open-source, with the original source and a short summary.
The latest AI and developer news about #open-source, with the original source and a short summary.
Kanzhun released the 33B dense model Nanbeige4.3-33B under the Apache 2.0 license.
The built-in provider supports Supervised and Auto modes using existing fx models and skills.
The model surpasses leading open-weight models on MindCube spatial reasoning and outperforms other world models on FlowBench motion prediction.
The release includes full model weights and a creation Skill, with INT4 and INT8 weights planned for later.
The framework integrates model serving, trajectory generation, and policy updates into a single reproducible pipeline supporting Qwen and GLM models.
The upgraded 560B-parameter model features a 1M-token context window and improved performance on reasoning and automation benchmarks.
The number of tasks where Qwen3.8-27B achieved 20 consecutive successes increased from 38 to 107.
StepFun released Step Code, derived from Pi, with built-in sub-agent and web deployment capabilities.
The library tokenizes PII before model calls and restores it only for authorized tools, supporting Spring AI 2.0.0 and Spring Boot 4.0.0.
The open-source tool rewrites its own prompts and compiled core through four verification gates, achieving a 4.3-second mid-conversation swap while maintaining 87.9% prompt cache retention.
The MIT-licensed tool transforms high-level ideas into Jira epics, vertical slices, and Cucumber tests, then executes the build and rollout via stacked PRs.
The open-source tool runs locally via CLI, browser, or desktop app without cloud calls, supporting PDF merge, split, OCR, and compression.
StepFun released the weights for its 600B MoE model Step 5 Preview and confirmed it can run on 8 H200s.
The 33B model serves as a teacher for upcoming smaller agents and outperforms Qwen3.8-27B on Terminal-Bench 2.1 but trails on SWE-Bench Pro.
The open-source agentic IDE features isolated git worktree workspaces, living specs, and a lightweight architecture that runs the pi agent in-process.
The MIT-licensed platform integrates seven AI agents for threat detection, 27 automated security actions, and continuous penetration testing across major cloud and security providers.
A training-free method using five coordinated model instances raises open-weight Qwen3.8-27B accuracy from 66.8% to 92.4% on hard coding tasks.
Destined AI releases Hurston-1, a 3B parameter Apache-2.0 model fine-tuned from Maya1 to generate accented English speech with native telephone-band output.
The free Python package ports models like Cloudflare's Clef and Perplexity's pplx-decider to MLX, achieving sub-20ms latency on small encoders on an M5 Pro and 96.4% accuracy on TREC after fine-tuning.
The Apache-2.0 licensed tool runs the pi coding agent in-process across isolated workspaces, available for macOS, Windows, and Linux.
The free and open-source tool allows users to download models from Hugging Face and access a local OpenAI-compatible API via Alt+Space.
The free beta tool compiles repositories into a queryable map to reduce source code reading by 2-6x compared to mainstream agents.
The 9B open-weights model achieves a 65 score on the AA Decision Index, matching d1-ultra while costing approximately $23 less per 1M tasks.
Modular has open-sourced Mojo 1.0 and released the source code for the MAX framework, while completing its acquisition by Qualcomm.