LiliPixel v1 is a 190M-parameter DiT image generator for high-quality pixel art, trained from scratch over 126 hours on an RTX 4080 and released under a fully open license with diffusers-ready weights on Hugging Face.
A teaser hinting that a new Olmo model training is underway, with a link for more details.
TontaubeAI released TontaubeV1-Verbalizer, an Apache 2.0 open-source LLM-based verbalization model fine-tuned from a Qwen3 1.7B backbone that converts written English text into spoken-form text for small TTS models, processing a 20,000-character document in under 2 seconds on an RTX 3090.
Reverb-TTS is an open-source zero-shot voice cloning TTS model that runs fully locally in the browser on WebGPU via ruNNtime, achieving about 17x realtime on an M4 Pro with no server round-trips.
Gemini 4 introduces fast, low-cost spatial reasoning for robotics, and can be paired with the new Robotics plugin in Google Antigravity to quickly build a working robot app.
KittenML released Kitten-TTS-2 Pico (7M/4MB), Nano (17M), and Mini (80M), a family of super-tiny multilingual open-source TTS models under Apache 2.0 that run faster than realtime on a single smartphone CPU thread and support up to 16 languages.
The author shares a first-hand experience using LFM3-1.7B, a small 1.7B parameter model, for medical long-context tasks (128k token patient notes summarization and QA), finding it surprisingly faithful, fast (10k tokens/s prefill on a 3090ti), and with minimal hallucination compared to much larger models like Qwen27B.
Google DeepMind 发布 Gemini 4 Neon,虽在榜单与定价上表现亮眼,但文章指出同一模型权重在更换 harness 后 SWE-sweep 成绩从 12.8% 跃升至 39.0%,引发对‘基准测试提升究竟来自模型还是工具链工程’的质疑。
Release of a ~320MB LoRA adapter RL-trained on Qwen3.8-27B inside Hermes Agent, achieving thinking-level accuracy with thinking disabled, 3.5–4.4× fewer tokens, and strong SWE-bench Pro / Terminal-Bench results, with GGUF weights available for local use.
OVH released the first model of its new model family, OVHaiLLM/Olala-v1.0-mini-7A1B, on HuggingFace.
OVHcloud's OVHai LLM releases Olala-v1.0-mini-7A1B, an open-weights 6.8B (1B active) hybrid Mamba-3/attention mixture-of-experts reasoning model trained on European supercomputers with 12.8T tokens, supporting chat, tool use, and four reasoning-effort levels under Apache 2.0.
A developer shares hands-on experience using Hugging Face's pi adapter LoRA on Qwen3.8 27B (Q4_K_S quant) for Penpot design workflows, reporting better tool/harness usage but slower token generation, occasional loop failures, and a lack of ready-made quants making it feel like a research preview rather than a production upgrade.
Cloudflare released the weights of Clairic and Clairic-flash, its document data-extraction models (27B and 9B parameters), under the Apache 2.0 license, enabling local parsing of invoices, receipts, and PDFs.
Gemini 4 Neon shows strong browser-use capability, scoring 74% — matching Opus 5.5 at half the price and up from 53% on 3.8 Flash in September — though still behind GPT-6.2 Sol's 82%.
Gemini 4 Neon is announced as a frontier reasoning model focused on faster task completion.
Sundar Pichai highlights Genie 4's ability to accept new prompts while exploring a generated world — adding characters or changing scenes with native audio adapting in real time — now available in Project Genie for eligible Google AI Ultra subscribers.
微软在 Ignite 大会上发布了七款自研 MAI 模型,覆盖推理、工具使用和长程任务等智能体常见工作负载,标志着微软加速构建第一方模型而非仅转售合作伙伴模型。
The author demonstrates converting their digital painting into a walkable 3D scene using Fable 5.2: the overall atmosphere, sky gradient, and color palette are quite close to the original, but the paths and details still look rough.
Google claims Gemini 4 Neon tops a simulated vending machine business benchmark while pricing at $3 per million tokens, with the author noting that lower per-task cost matters more than leaderboards for what developers actually build.
Gemini 4 Neon is positioned as a competitor to Anthropic's Opus 5.5, signaling intensified rivalry among frontier AI model providers.