DeepReinforce Releases Self-Improving Coding Model "Ornith-1.0"

DeepReinforce AI has released Ornith-1.0, an open-source model family specialized for agentic coding. It comes in four sizes — 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE — available under the MIT license.

Its standout feature is a "self-improving training framework": through reinforcement learning, it simultaneously optimizes not only the solutions but also the scaffolds that lead to them. According to the team, this enables the model to autonomously discover better exploration trajectories and generate high-quality solutions.

On benchmarks, it posts scores that significantly outperform existing open-source models in the same size class on Terminal-Bench 2.1, SWE-Bench Verified, and others. Notably, the 35B-MoE model achieves 75.6% on SWE-Bench Verified, closing in on Qwen3.5-397B (76.4%). Built on Gemma 4 and Qwen 3.5 with post-training, it is also practical as a lightweight model that runs on a single GPU.

LiquidAI Announces "LFM2.5-230M", a 230M-Parameter Edge Model

LiquidAI has released LFM2.5-230M, a hybrid model designed for on-device deployment. Pretrained on 19T tokens, it achieves decoding speeds of 213 tok/s on the Galaxy S25 Ultra and 42 tok/s even on a Raspberry Pi 5.

The LFM2.5 architecture features a 14-layer structure (8 double-gated LIV convolution blocks + 6 GQA blocks) and supports a context length of 32,768 tokens. It supports 10 languages, including Japanese. It also comes with tool-use capabilities, with intended applications in data extraction and lightweight agent pipelines.

However, it is not recommended for reasoning-heavy tasks (advanced mathematics, code generation, creative writing); it is positioned squarely as a model dedicated to practical task processing on edge devices.

Anthropic Accuses Alibaba of "Largest-Scale AI Distillation Attack"

Anthropic has accused Alibaba of conducting a massive AI distillation attack against its models. The attack allegedly involved approximately 28.8 million unauthorized API requests, believed to be the largest of its kind in the industry.

A distillation attack is a technique in which large volumes of requests are sent to a powerful commercial model and the responses are used to train the attacker's own models. It very likely violates API terms of service and is also considered problematic from an intellectual property perspective. Whether this case leads to tighter API usage restrictions or stronger legal measures going forward is drawing close attention.

Norway to Ban AI Use in Primary Schools

Norwegian Prime Minister Jonas Gahr Støre has signaled a policy to ban AI use in primary schools. The move is driven by concerns that using AI increases the risk of children missing critical steps in their education. Norway previously banned smartphones in schools as well, and it maintains a cautious stance toward introducing digital technology into the classroom.

Oracle's Headcount Down About 21,000 Amid AI Adoption

It has emerged that Oracle's workforce has shrunk by about 21,000 employees, with AI-driven efficiency gains seen as the likely cause. The argument that "AI takes tasks, not jobs" has also been debated on Hacker News, and it is fair to say we are beginning to see a structure in which parts of business processes get automated — ultimately leading to headcount reductions — rather than entire job categories disappearing outright.

Linux Foundation Launches Open Source Defense Project "Akrites"

The Linux Foundation has announced Akrites, a new project to defend FOSS (Free and Open Source Software) against AI-powered exploits. As cyberattacks abusing AI grow more sophisticated and more numerous, securing open source software has become an urgent priority. Concrete results are still to come, but the significance of a major foundation demonstrating an organized commitment to this problem is substantial.

audio.cpp: C++ Audio Model Inference Framework Up to 5x Faster Than Python

audio.cpp, an open-source C++ audio model inference framework, is drawing attention in Reddit's LocalLLaMA community. Built on ggml, it can run 12 models — spanning TTS, ASR, voice conversion, and codecs — in a single C++ runtime.

Compared with Python-based reference implementations, it achieves speedups of up to 3.68x on PocketTTS, up to 2.74x on Qwen3-TTS, and 5.03x on Vevo2. On long-form text, PocketTTS generates speech at 48.40x real time, making the framework a promising option for bringing audio AI to edge environments.

---