1. Meta Releases Muse Glimmer, a 30B Open-Weight Agentic Model
Meta has returned to an open-weight AI strategy with the release of Muse Glimmer, its first open-weights model since Llama 4. Distilled from the larger Muse Spark model, Muse Glimmer is designed specifically for local, always-on agent workflows like end-to-end task completion and tool calling. The model is compatible with major local runtimes including Ollama, LM Studio, vLLM, and SGLang, and is available for download on Hugging Face.
- • Muse Glimmer is a 30-billion-parameter dense causal transformer featuring a 1.8-billion-parameter perception encoder for multimodal text and image inputs.
- • The model is released under the permissive Apache 2.0 license, allowing unrestricted commercial use, modification, and redistribution.
- • It supports a context length of over 128,000 tokens and is optimized to run on consumer GPUs with 24GB to 32GB of VRAM using 4-bit quantization.
- • The release includes DFlash speculative decoding, which Meta reports delivers up to a 3.1x generation speedup on an Nvidia RTX 5090.
- • Meta announced plans to release the weights for its more powerful Muse Spark 1.2 model in the near future.
Developers can now run a highly capable, permissive open-weight agentic model locally on consumer hardware without relying on cloud APIs.
2. Cactus Launches Needle 2, Successor to Original Edge LLM
Building on the original 26M parameter Cactus Needle released in May 2026, Cactus has launched Needle 2. This new iteration is a 45-million-parameter model compressed to 2 bits, enabling it to run on just 28MB of RAM. While the original model focused on general local execution, Needle 2 is specifically optimized for agentic tasks like structured extraction and tool-calling on edge hardware, achieving significantly higher token throughput on devices like the Raspberry Pi 5 and Meta Quest 3S.
- • Needle 2 is a 45M parameter model compressed to 2 bits, requiring 28MB of RAM.
- • It improves on the original Cactus Needle by adding native support for structured extraction and tool-calling.
- • The model achieves 500 tokens/sec on Raspberry Pi 5 and 1,500 tokens/sec on Meta Quest 3S.
- • Includes a learned confidence score to manage local vs. cloud execution.
- • Uses Simple Attention Networks with 70 MFLOPs per token efficiency.
This release advances the Cactus Needle line by enabling more complex agentic workflows on ultra-constrained hardware, moving beyond the capabilities of the initial 26M parameter release.
3. Anthropic Eliminates Classifier Fees for Claude Code
Building on the recent announcement that auto mode will become the default setting for Claude Code on August 14, 2026, Anthropic has further updated the service by eliminating all classifier overhead fees. This change, effective immediately, removes the additional costs previously associated with the safety classifier that intercepts potentially dangerous commands.
- • Anthropic has eliminated all classifier overhead charges for Claude Code users on paid plans, effective immediately.
- • This follows the previously announced transition to auto mode as the default setting for new sessions starting August 14, 2026.
- • The classifier remains active to block destructive actions and trigger manual approval when necessary.
- • New safety features include hard denies for data exfiltration and pre-action git status checks.
The removal of classifier fees lowers the cost of running autonomous agent workflows, making the default auto mode more accessible for developers on Pro, Max, and Team plans.
4. Ante Launches as a 15MB Offline Coding Agent in Rust
Ante has launched in alpha preview as a lightweight, high-efficiency alternative to terminal coding tools like Claude Code. Written in Rust, the single binary manages its own local llama.cpp instance to execute GGUF models offline, significantly reducing memory, CPU, and disk I/O usage. Telemetry is enabled by default but can be completely disabled by setting the ANTE_TELEMETRY environment variable to off.
- • Ante is a ~15MB Rust binary coding agent for macOS and Linux that supports fully offline operation.
- • It features a built-in, managed version of llama.cpp to run GGUF models locally without API keys or internet access.
- • The agent supports over 12 LLM providers and allows custom provider configuration via environment variables or a JSON catalog.
- • It reports an 82.7% success rate on Terminal-Bench 2.1 when paired with the DeepSeek V4 Flash model.
- • The SDK, protocol, and evaluation pipeline are licensed under Apache 2.0, while the core harness is distributed as a prebuilt binary.
Developers can run a highly efficient, fully offline coding agent on local hardware without incurring API costs or sending code to external servers.
5. Spotify Open-Sources Xirp Coding Agent Orchestrator
Spotify has opened the public beta of Xirp, an internal developer tool designed to orchestrate multiple AI coding agents. Xirp enables developers to run and compare agents like Claude Code, Gemini CLI, and OpenAI Codex side-by-side, leveraging workflows refined over tens of thousands of internal sessions to streamline agentic development.
- • Spotify has released its internal tool, Xirp, to the public in beta.
- • Xirp allows users to run Claude Code, Gemini CLI, and OpenAI Codex agents side-by-side.
- • The tool has been battle-tested internally across more than 36,000 coding-agent sessions at Spotify.
Developers can compare, run, and orchestrate multiple frontier coding agents simultaneously within a single unified interface.
6. Juror Offers Open-Source PR Reviews Using Parallel LLMs
Juror has launched as an open-source alternative to commercial PR review tools like Greptile. Designed to run as a GitHub Action or CLI tool, Juror coordinates a "jury" of different LLMs to analyze code diffs, using code-aware similarity and a referee model to collapse duplicate reports. The tool prioritizes security by isolating credentials and enforcing read-only access, ensuring that code never leaves the local runner environment except via direct API calls to the configured model providers.
- • Juror runs multiple frontier models in parallel to review GitHub pull requests and deduplicates findings using a referee model.
- • The tool runs entirely on the user's own GitHub Actions runner, keeping code secure except for model API calls.
- • It supports multiple API providers including OpenAI, Anthropic, xAI, and Fireworks, with customizable jury presets.
- • Security features include per-harness credential isolation, read-only repository access, and base-revision policies to prevent configuration tampering.
- • Each review generates a cost receipt detailing the models used, time taken, and estimated spend.
Developers can deploy a secure, self-hosted code review assistant that leverages multiple LLMs without exposing source code to third-party SaaS platforms.
7. AWS Continuum Integrates Claude Code and Codex for Automated Vulnerability Remediation
Building on the prior availability of OpenAI's Codex on Bedrock and Anthropic's Claude Code security plugin, AWS has integrated both tools into its Continuum platform. Announced at Black Hat USA 2026, this integration allows developers to use an agent-team loop to automate the lifecycle of vulnerability remediation, alongside new GuardDuty monitoring for AI-related cost harvesting and a free AI inventory tool in Security Hub.
- • AWS Continuum now orchestrates Claude Code and OpenAI Codex in an agent-team loop for vulnerability remediation.
- • The platform automates the discovery, prioritization, validation, and remediation of code vulnerabilities.
- • AWS GuardDuty now monitors data plane events to detect unauthorized AI inference (cost harvesting).
- • AWS Security Hub added a supply chain protection category and a free AI inventory capability.
This integration moves beyond standalone tool usage by providing a unified, automated remediation loop within the AWS cloud environment.
8. Anthropic Cancels Planned Price Hike for Claude 3.5 Sonnet
Anthropic has announced that the introductory pricing for Claude 3.5 Sonnet, previously set to expire on August 31, 2026, will now remain in effect permanently. This update overrides the earlier plan to increase rates to $3 per million input and $15 per million output tokens, maintaining the current $2/$10 rate for developers.
- • Introductory pricing of $2 per million input and $10 per million output tokens is now permanent.
- • The previously scheduled price increase to $3/$15 on September 1, 2026, has been cancelled.
- • This change ensures long-term cost stability for applications built on Claude 3.5 Sonnet.
Developers can now rely on the current, lower pricing for Claude 3.5 Sonnet indefinitely, removing the need to adjust budget projections for the previously planned September price increase.