Three open-weight model releases landed in the last week that matter for different reasons. Z.ai shipped GLM-5.3 with a coding focus, Xiaohongshu's Dots studio released its first public dots3 model, and Meta open-sourced Muse Glimmer 30B, a model that runs on a single consumer GPU. Each one makes a different argument about where open-weight AI is heading.
GLM-5.3: post-training does the heavy lifting
Z.ai, the international brand of Zhipu, released GLM-5.3 on top of the same base model as GLM-5.2. All the gains come from post-training. The company reports a 50% improvement on its internal Z.ai Code Bench, plus leading open-source results on Terminal-Bench 3.0 and Agents' Last Exam.
The interesting part is the cybersecurity angle. Z.ai says GLM-5.3 shows stronger vulnerability-discovery and exploitation performance. That cuts both ways. A model that is better at finding vulnerabilities in a sandbox is also better at finding them in production systems. The weights ship two weeks after launch following safety evaluation, which is at least an acknowledgement that the problem exists.
One thing worth noting: the 50% number is against Z.ai's own benchmark. Third-party results will tell the real story. Still, getting that kind of gain from post-training alone, without touching the base model, is a sign that the industry's focus on RL and training-time compute is paying off.
dots3-note: a 512K context MoE under Apache 2.0
Xiaohongshu's Dots studio released dots3-note Preview on 14 August 2026. It is a 280 billion parameter mixture-of-experts model with 16 billion active parameters. Context window is 512,000 tokens. Input accepts text, images, video, and audio. The license is Apache 2.0, confirmed in the license file, so commercial use is permitted.
The model architecture separates perception from language. Vision runs through a 7 billion parameter MoE vision transformer with 1.2 billion active parameters. Audio goes through a dense 800 million parameter encoder. If you only need the language model, both vLLM and SGLang expose a flag to load it without the perception encoders. That is a practical detail that matters for teams that want to run this on modest hardware.
The release also introduces TEMPO, a reinforcement learning method Dots built to train agents on tasks that take tens of hours. Most RL for agents gets trained on tasks that finish in seconds or minutes. Ten-hour horizons change what the model learns. It has to hold a goal, plan, execute, recover from failures, and only then get a reward. Few open-source projects have published methods for that scale of training.
Weights are on Hugging Face as dots-studio/dots3-note-prev and on ModelScope under the same names. Code is on GitHub at studio-dots-ai/dots3-note-prev. OpenRouter also carries it for free as dots-studio/dots-3-note-preview with the full 512K context.
Muse Glimmer: local agents without the API bill
Meta's Muse Glimmer 30B is distilled from Muse Spark and optimized for always-on local agent workflows. The weights are Apache 2.0. It runs on a single consumer GPU or a Mac without any network calls.
Architecture specifics: dense causal Transformer with grouped-query attention using 32 query heads and 2 KV heads. Attention pattern is local-local-local-global with a sliding window of 2048. The vision encoder is a ViT-G/14 with about 1.8 billion parameters, supporting up to 4,096 visual tokens per image. Context length is 131,072 tokens. Vocabulary is 202,048. Knowledge cutoff is 4 January 2026.
Full precision needs over 55GB of memory, which is why Meta released two quantized versions. K-Quant-Dynamic targets 32GB VRAM with an average degradation of 0.2%. K-Quant-17GB targets 24GB VRAM with an average degradation of 1.0%. Block-level speculative decoding improves speed. I agree with Meta's decision to ship these quantized versions as first-class artifacts instead of treating quantization as an afterthought. For people running agents locally, the 17GB version is the one that matters.
The license is more permissive than previous Llama licenses. Simon Willison tested the 18.16GB quantized version on his blog for coding agents and image description, and confirmed it runs on machines with 32GB+ RAM alongside other applications. That is the exact use case: a model that lives in the background and can be invoked without waking up the cloud.
What these three releases have in common
All three are Apache 2.0 or permissively licensed. All three are designed for agentic workloads. GLM-5.3 targets coding and terminal tasks. dots3-note has long-context and long-horizon agent training. Muse Glimmer is explicitly built for always-on local agents.
The other pattern is the focus on post-training and distillation rather than base model innovation. GLM-5.3 reuses the GLM-5.2 base. Muse Glimmer is distilled from Muse Spark. The marginal cost of pushing a base model further is getting higher, a reality that is pushing the industry toward the more bottom-line-relevant parts of the stack: how well the model follows instructions, how long it can hold a goal, how fast it can run on hardware people already own.
Practical takeaways
For teams building agents, the dots3-note Preview is worth a serious look. A 512K token context under Apache 2.0 with free serving on OpenRouter removes two of the usual blockers: cost and license friction. The TEMPO method is also documented, which is a better source of long-horizon agent training learnings than any benchmark leaderboard.
For local deployments, Muse Glimmer is the first open-weight model I would actually consider putting in an agent loop on a laptop. The 17GB quantized version runs on a 24GB GPU or an M4/M5 Max Mac, and the 0.2% degradation on the 32GB version is not a meaningful downside for most agent tasks. The speculative decoding setup is already integrated, so throughput is not the usual quantized-model mess.
For security teams, GLM-5.3's vulnerability-discovery results deserve cautious attention. A model that can reliably find and exploit vulnerabilities will change how penetration testing works, and it will also change how much damage a misconfigured exposure can cause. Z.ai's decision to hold the weights for two weeks of safety evaluation is reasonable, but two weeks is not forever, and once the weights are out, they are out.
I genuinely do not know how to feel about the security implications here. The capability is real and the release is responsible within the current norms, but the norms themselves may not be adequate for what these models can do. That is not a criticism of Z.ai specifically. The whole field has been moving faster than evaluation processes for two years now.
What I am watching next
The dots3-note weights just shipped, so third-party evaluation of the long-context claims is the first thing to check. A 512K context is aggressive, and MoE models often drop quality in the middle of the context window. Someone needs to run the needle-in-a-haystack and multi-document retrieval benchmarks with real data, not the paper's numbers.
For Muse Glimmer, the integration story matters. Optimized llama.cpp, MLX, and ExecuTorch builds are coming soon, but "coming soon" is not "available now." The DFlash drafter is already out, which helps. If the ExecuTorch build materializes in a reasonable state, this becomes the default local agent model for Mac users.
The broader trend is that open-weight models are no longer merely fine for prototyping. They are becoming the actual production deployment for a class of workloads: local agents, long-horizon tasks, and anything where per-token API costs would make the product impossible. The prices for hosted APIs are not going to keep falling forever. A model that runs on the hardware you already have is a hedge against that.