open weights
30 articles about open weights in AI news
Alibaba Qwen3.8-Max Open Weights Due Next Week
Alibaba announced Qwen3.8-Max, its most capable model, with open weights due next week. No specs disclosed yet; release will test open-weight competitiveness.
Minimax to Release Open Weights in Two Weeks, Highlighting Chinese Startup Momentum
Chinese AI startup Minimax announced it will release open weights within two weeks. This follows a pattern of rapid open-source releases from Chinese firms, contrasting with Meta's more controlled approach.
DeepSeek V4-Pro: 1.6T parameters, open weights, undercuts rivals 10x
DeepSeek unveiled V4-Pro and V4-Flash, its largest open-weight models with up to 1.6 trillion parameters and a 1M-token context window. The new hybrid attention architecture cuts compute for long contexts by 73–90%, enabling prices far below OpenAI, Google, and Anthropic.
Kimi 2.6 Thinking Shows Promise as Open Weights Model, Lags Behind Closed SoTA
An initial evaluation of Moonshot AI's Kimi 2.6 Thinking model finds it generates extensive reasoning traces but delivers only 'okay-ish' results on creative and coding tasks, highlighting the persistent open vs. closed model gap.
Meta's 'Spark' AI Model Leaked as Closed-Source, Breaking Open-Weight Streak
A leak suggests Meta's new 'Spark' AI model will not be released with open weights, marking a significant departure from its strategy of open-sourcing foundational models like Llama.
Redis Creator antirez Writes MiniMax H3 Mac Inference Engine
Redis creator antirez wrote a MiniMax H3 Mac inference engine, announced on X. Open weights enabled the port, highlighting community-driven hardware support.
MiniMax H3 Video Model Beats Seedance 2.0, Opens Weights
MiniMax launched H3 video model, ranking #1 in editing benchmarks while opening weights to challenge ByteDance's Seedance 2.0 and Google's Gemini Omni Flash.
Qwen 3.6 Released: Free, Open-Weights Model for Local AI Coding
Alibaba's Qwen team released Qwen 3.6, an open-weights AI model for local deployment. This provides a free, private alternative to ID-verified models like Anthropic's Mythos and OpenAI's Codex.
Google DeepMind: Web Environment, Not Model Weights, Is Key AI Agent Attack Surface
Google DeepMind researchers present a systematic framework showing that the web environment itself—not just the model—is a primary attack surface for AI agents. In benchmarks, hidden prompt injections hijacked agents in up to 86% of scenarios, with memory poisoning attacks exceeding 80% success.
MiniMax-M3 Scores 55 on AI Index, Open-Source Lead Looms
MiniMax-M3 scored 55 on the Artificial Analysis Intelligence Index, set to become the leading open-source model once weights are released.
Google Releases Magenta RealTime 2 for Open-Weight Music Generation
Google released Magenta RealTime 2 on Hugging Face, the only open-weights model for real-time continuous music generation on device with ~200ms latency.
Miso One: 8B Open-Source TTS Hits 110ms Latency, Real Emotion
Miso One, an 8B open-source TTS model, achieves 110ms latency with emotional range. Weights are fully open-source for self-hosting, but no benchmark data is provided.
Mistral AI Launches Voxtral TTS: 3B-Parameter Open-Source Model Claims 63% Win Rate Over ElevenLabs Flash v2.5
Mistral AI released Voxtral TTS, a 3-billion-parameter open-weights text-to-speech model. It reportedly outperforms ElevenLabs Flash v2.5 in human preference tests, runs on 3 GB RAM, and clones voices from 5 seconds of audio.
AI2's MolmoWeb: Open 8B-Parameter Web Agent Navigates Using Screenshots, Challenges Proprietary Systems
The Allen Institute for AI released MolmoWeb, a fully open web agent that operates websites using only screenshots. The 8B-parameter model outperforms other open models and approaches proprietary performance, with all training data and weights publicly released.
Chinese Lab's Free MoE Model Matches GPT-5.5 on Agentic Coding
A Chinese lab released an Apache-2.0 open-weights MoE model matching GPT-5.5 on agentic coding. This free model challenges proprietary AI's lead with sparse MoE architecture.
GPT-5.4 Pro Solves 60-Year-Old Erdős Problem #1196, Finds 'Book Proof'
OpenAI's GPT-5.4 Pro solved Erdős Problem #1196, a 60-year-old conjecture on primitive sets, in ~80 minutes. The AI discovered a purely analytic proof using von Mangoldt weights, rejecting the standard probabilistic approach used by mathematicians since 1935.
Nvidia Trains Billion-Parameter LLM Without Backpropagation
Nvidia demonstrated training a billion-parameter language model using zero gradients or backpropagation, eliminating FP32 weights entirely. This could dramatically reduce memory and compute costs for LLM training.
LLM-HYPER: A Training-Free Framework for Cold-Start Ad CTR Prediction
A new arXiv paper introduces LLM-HYPER, a framework that treats large language models as hypernetworks to generate parameters for click-through rate estimators in a training-free manner. It uses multimodal ad content and few-shot prompting to infer feature weights, drastically reducing the cold-start period for new promotional ads and has been deployed on a major U.S. e-commerce platform.
Kimi 2.5's 1T Parameter MoE Model Runs on 96GB Mac Hardware via SSD Streaming
Developers have demonstrated that Kimi 2.5's 1 trillion parameter Mixture-of-Experts model can run on Mac hardware with just 96GB RAM by streaming expert weights from SSD, with only 32B parameters active per token.
Qwen 3.5 397B-A17B MoE Model Runs on M3 Mac at 5.7 TPS with 5.5GB Active Memory via SSD Streaming
Developer Dan reportedly runs the 209GB Qwen 3.5 397B-A17B MoE model on an M3 Mac at ~5.7 tokens per second using only 5.5GB of active memory by quantizing and streaming weights from SSD.
MetaClaw: AI Agents That Learn From Failure in Real-Time
MetaClaw introduces a breakthrough where AI agents update their actual model weights after every failed interaction, moving beyond prompt engineering to genuine on-the-fly learning without datasets or code changes.
Qwen3.8-2.4T-A95B Open-Weight MoE Hits Hugging Face
Qwen3.8-2.4T-A95B, a 2.4T-parameter open-weight MoE with 95B active parameters, was released on Hugging Face, marking a new scale for open-weight models.
MSS Officer Gets Root on OpenAI Cluster via Irregular Sandboxes
MSS officer allegedly gained root on OpenAI cluster via irregular sandboxes, per @teortaxestex. Unverified claim highlights AI infrastructure security risks.
NVIDIA Releases Nemotron VoiceChat, First Open Full-Duplex Speech Model
NVIDIA released Nemotron VoiceChat, claiming the first open full-duplex speech model with tool calling and barge-in. The move targets real-time voice agents, challenging proprietary APIs.
DeepSeek-V4-Flash Open-Sourced: 304B Model Beats V4-Pro at $0.14
DeepSeek open-sourced V4-Flash-0731, a 304B model scoring 82.7 on VulcanBench, matching Claude Opus-4.8 at $0.14/M input tokens.
Open-Weight Models Just Matched Claude Opus 4.6 — Here's How to Run Them
Route Claude Code to open-weight models like Kimi K3 and DeepSeek V4 Flash via ANTHROPIC_BASE_URL or LiteLLM. Test cheap models before spending Opus 4.6 credits. The open-weight revolution is now a Claude Code workflow decision.
Nvidia, Meta, Mistral Warn US Against Broad Open-Weight AI Restrictions
Nvidia, Meta, Mistral sign letter urging US against broad open-weight AI restrictions, arguing distillation is legitimate. Comes as White House weighs response to Chinese AI model distillation.
Microsoft Fara1.5-27B Open-Source Agent Scores 72.3% on Web Tasks
Microsoft released Fara1.5-27B, a vision-only web browsing agent scoring 72.3% on Online-Mind2Web, open-source on Hugging Face.
Alibaba Qwen3.8: 2.4T Parameter Open-Weight Model Incoming
Alibaba's Qwen3.8, a 2.4T parameter open-weight model, was announced. It would be the largest open-weight model ever, but lacks benchmark details.
Murati's Thinking Machines Ships 975B Inkling — Leads US Open Models
Murati's Thinking Machines releases Inkling, a 975B-parameter MoE model that leads US open models but trails Chinese rivals on benchmarks and cost.