Tokens by Michael Brenndoerfer, Week #2
NVIDIA's open agent brain gets a giant context window
Today's issue: A cathedral-sized context window, a terminal-shaped Colab, and a chatbot VTuber trying to maintain eye contact.
Welcome to #2 of your weekly AI newsletter.
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
Main Story
NVIDIA's Nemotron 3 Ultra is built for long-running agents
Every time I give an agent a bigger context window, I tell myself it will use the extra room responsibly. Then it reads 900 log files, forms a strong opinion about my naming conventions, and suggests a quick cleanup that somehow touches auth.
NVIDIA released Nemotron 3 Ultra, a 550B total-parameter, 55B active-parameter open Mixture-of-Experts model aimed at long-running autonomous agent tasks. It uses a hybrid Mamba-Attention architecture, was pretrained on 20T text tokens, and was extended to a 1M-token context length.
For users, the useful part is speed and control. In NVIDIA's tests, the model reaches up to roughly 6x higher inference throughput than publicly available LLMs at on-par accuracy, and reasoning effort control lets users spend more compute when they want better answers.
NVIDIA is also releasing Base BF16, Post-Trained BF16, NVFP4 quantized, and GenRM checkpoints, plus training recipes, data, and RL environments. An open model is more useful when it ships with the machinery to reproduce and adapt it, not a giant zip file and a README shaped like a treasure map.
- Agent-shaped architecture: Nemotron 3 Ultra is built for long-running autonomous tasks, with a hybrid Mamba-Attention design and a 550B total-parameter MoE setup where 55B parameters are active per pass.
- Giant working memory: The model supports a 1M-token context window, enough room for agents juggling code, logs, tool outputs, and the ancient README nobody updated.
- Throughput focus: NVIDIA reports up to roughly 6x higher inference throughput than comparable publicly available LLMs in its tested settings, helped by Multi Token Prediction for speculative decoding.
- Open release package: The release includes multiple checkpoints, training recipes, data, RL environments, and a refreshed 173B-token GitHub code dataset with a September 30, 2025 cutoff.
Bottom Line: Nemotron 3 Ultra targets the less glamorous part of agent hype: making long-horizon systems cheaper and more practical to run on NVIDIA's stack. The open checkpoints and recipes make it interesting for researchers and infra teams, but the gains still have to show up in agents that finish useful work without turning every repo into a group project.
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
The Feed
-
PYMNTS cites FT on a reported ChatGPT overhaul toward Codex, agents, images, partner apps, and a cross-device personal agent. "Chat is dead," says the deck.
-
Gemma 4 12B: Google DeepMind's open multimodal model feeds raw image patches and audio waveforms into the LLM backbone for lower-latency audio and visual reasoning on 16GB laptops.
-
Headroom compresses tool outputs, logs, files, and RAG chunks before they reach the model, claiming 60 to 95% fewer tokens with the same answers. My stack traces feel seen.
-
Google's Colab CLI brings remote CPU, GPU, and TPU runtimes to your terminal for scripts, notebooks, stdin, REPLs, and shells, then tears down the VM when you are done.
-
GrepSeek: UMass Amherst trained Qwen3.5-9B to answer by issuing Unix commands like
rg,grep, andheadover raw text, skipping vector indexes and their setup tax. -
Researchers from NTU, NUS, and CUHK shared Audio-Interaction, an open streaming audio model that listens continuously, decides when to respond, and includes a 2.6M-item instruction corpus.
-
Alibaba's Open Code Review pairs deterministic checks with LLM review in a CLI, with GitHub Actions, Claude Code integration, and checks for NPEs, thread safety, XSS, and SQL injection.
-
LongTraceRL: THU-KEG trains long-context reasoning with search-trajectory distractors and entity-level rubric rewards, basically grading the model's scratch work during the open-book exam.
-
Matt Van Horn's last30days-skill researches topics across Reddit, X, YouTube, Hacker News, and the web, then writes grounded briefs. With 34.5k stars, the research gremlin is popular.
-
Aspire 2B is a new Singapore research supercomputer with the country's largest advanced Nvidia GPU cluster. It gives 9,000-plus public researchers local compute for weather, disease, and language AI.
-
Apple introduced Siri AI. The crawl only confirms Apple's headline, so I am not handing Siri imaginary powers today.
-
Domino: this speculative decoding implementation keeps draft generation block-parallel, then adds a lightweight causal correction head to improve draft-token acceptance. Guess first, verify smarter.
-
Open-LLM-VTuber gives "any LLM" local hands-free voice chat, interruption, and a Live2D talking face. It has 10.5k stars, so my chatbot now has eye contact before ignoring instructions.
-
dot-skill: this 19.2k-star repo packages "colleague skills" for agents with AgentSkills badges and files like
SKILL.md. My next teammate may arrive as Markdown with opinions. -
SJTU researchers proposed Representation Forcing, which trains multimodal models to predict visual representations before pixels. It removes the frozen VAE bottleneck while matching strong image-generation baselines.
See you next week!
Michael
Get Tokens in your inbox 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
Join 5,000+ readers · No spam, unsubscribe anytime.