A coding AI you can run and modify yourself

Meta released Muse Glimmer, a 30-billion-parameter model for local agents, under the Apache 2.0 license. It accepts text and images and targets coding and tool use. Meta says compression brings the language-model weights from more than 55 GB to under 20 GB, with deployment possible in 24 GB or 32 GB including working memory and auxiliary components.

Ollama added support, including image input through its Apple Silicon MLX engine. Its announcement lists document reading, screenshot-driven computer use, and building applications from mockups as uses. Ollama also lists a context window exceeding 128,000 tokens and selectable reasoning effort, so users can adjust how much computation the model spends on a task.

Top Stories#

Anthropic reports a mathematical advance related to the Riemann hypothesis. The hypothesis concerns the nontrivial zeros of a function describing the distribution of prime numbers. It predicts where those zeros lie; the full problem remains unsolved.

Anthropic says Claude raised the lower bound on the proportion lying on the predicted line from 41.6% to 67.2%.

The model combined established research over two sessions totaling 31 million output tokens. Anthropic says its mathematicians examined the result and a Lean formalization passed its stated validation tool. Anthropic does not expect these techniques to prove the full hypothesis.

Researchers extracted reasoning from encrypted API responses. A new paper describes replaying encrypted reasoning blocks into weaker models from the same provider, then inducing those models to reveal the contents. Encryption alone did not make publicly shared agent logs safe in these experiments.

The authors describe extraction across Anthropic, OpenAI and Google, along with hidden prompt injection and leakage from reasoning even when visible answers refused a request. Their reproducibility statement says provider mitigations made the specified attacks no longer reproducible as of August 2026.

Raindrop released a classifier for behavior in agent logs. Signals 2.0 generates code to extract relevant evidence from production traces, then trains a specialized classifier.

That replaces repeatedly asking a large model to inspect the entire history. One example detects an agent claiming success after repeated tool failures.

In Raindrop's evaluation, four interpretations of one behavior matched 0.9%–4.6% of the same 2,000 traces. Changing the definition changed how often the monitor found the behavior, even with the logs held fixed.

Top Links#

• Daybreak on AWS (openai.com): OpenAI made its cybersecurity capabilities available through Amazon Bedrock, with enrollment and approval required for both defensive security and authorized vulnerability research offerings.
• Regional inference (mistral.ai): Mistral's generally available endpoints offer European or US processing, while expressly permitting limited, safeguarded transfers to subprocessors outside the selected region.
• Grok 4.6 (x.ai): SpaceXAI released the model for sustained agent work, coding, and interactive or visual projects, with access through Cursor and Grok Build.