👋 Jan
@jandotai
Jan is an open source ChatGPT-alternative that runs 100% offline. Built by @menloresearch. Community: https://discord.gg/TE5wMUa7b6
Qwen releases Qwen3-Coder, an open agentic coder that rivals Claude Sonnet 4. - 480B total params, 35B active (MoE) - 256K native context, extendable to 1M with YaRN - Excels at agentic coding and browser-based tool use
>>> Qwen3-Coder is here! ✅ We’re releasing Qwen3-Coder-480B-A35B-Instruct, our most powerful open agentic code model to date. This 480B-parameter Mixture-of-Experts model (35B active) natively supports 256K context and scales to 1M context with extrapolation. It achieves…
Introducing Lucy: 1.7B model that Google for you It's an agentic‑search model that can even run on your phone. - Agentic search on tap - Lucy calls tools (<think></think>‑aware) - Fits in your pocket - runs on CPU or mobile Under the hood: - Built on @Alibaba_Qwen's Qwen3‑1.7B…
This is really interesting. Someone built an AI Airport Simulation to test how LLMs handle real-time, high-stakes decision making like fuel like fuel emergencies, collision avoidance, and traffic congestion. Open-source & stress-tested. github.com/jjasghar/ai-ai…
ByteDance releases Seed-X, a 7B multilingual translation model. It's trained with instructions, RL, and reward tuning. Reportedly outperforms Gemini 2.5, Claude 3.5, and GPT-4 on translation across 28 languages, based on human evals & automatic metrics. huggingface.co/ByteDance-Seed…
OmniSVG-3B weights are now available. A multimodal model for generating scalable vector graphics, icons, logos, even anime characters from text or images. Also includes a 2M-sample dataset (MMSVG-2M) and supports both local and online demos. huggingface.co/OmniSVG/OmniSVG
Qwen3-235B's latest version leads across 5 major reasoning and knowledge benchmarks. It's beating Claude Opus, GPT-4o, and DeepSeek V3 on GPQA, AIME25, LiveCodeBench, Arena-Hard, and BFCL.
Qwen announced Qwen3-235B-A22B-Instruct-2507, dominating math reasoning with 70.3 on AIME25 (vs GPT-4o's 26.7) and achieving 95.0 on ZebraLogic. It outperforms Opus 4 and GPT-4o across reasoning & coding while using 22B active parameters. huggingface.co/Qwen/Qwen3-235…
Audio search without transcription: ColQwen-Omni embeds 30 minutes of audio in 10 seconds across documents, audio, and video in a single 3B model. huggingface.co/vidore/colqwen…
After spending 12 hours on a plane yesterday, I was really missing the support of LLMs for research. Currently exploring jan.ai which seems to be a very nice and easy way to run local open source models on my oldish Mac m1 air. Have a feeling this endeavour might…
A new Russian-aligned 32B model just beat Qwen3-32B and DeepSeek R1 on ruMMLU, ruAIME, and LiveCodeBench. T-pro-it-2.0 builds on Qwen3 with continual pretraining, 500K instructions, and a toggleable reasoning mode. huggingface.co/t-tech/T-pro-i…
Pusa is a new video diffusion model that matches SOTA with 200x less training cost & 2500x less data. It outperforms Wan-I2V on VBench-I2V, runs 5x faster, and supports I2V, T2V, start-end frames, video extension, and video completion. huggingface.co/RaphaelLiu/Pus…
Jan v0.6.5 is out: SmolLM3-3B now run locally Highlights 💫 - Support for @huggingface's SmolLM3-3B - Fully responsive design across all screen sizes - New layout for Model Providers Update your Jan or download the latest.
250 tokens/sec with Kimi K2 in Jan via @GroqInc. To set it up: - Go to Settings: Model Providers: Groq - Paste your Groq API key - Add model: moonshotai/kimi-k2-instruct
Menlo Research is hiring, come build cool stuff with us. Some open roles: - Sr Backend Engineer (Tauri) - Sr Software Engineer - Research Engineer (PyTorch) - Lead Electromechanical Systems Engineer - Mechanical Engineer - Management Associate menlo.bamboohr.com/careers/
Microsoft releases a new dataset that improves Qwen2.5-7B from 17.4% to 57.3% on LiveCodeBench. It's called rStar-Coder, 418K tasks designed to push competitive code reasoning. A 7B model trained on it outperforms QWQ-32B on the USA Computing Olympiad. huggingface.co/datasets/micro…
