1. Google DeepMind Releases Gemini Robotics 2
Google DeepMind's Gemini Robotics 2 suite introduces three models: Gemini Robotics 2 (VLA), Gemini Robotics ER 2 (embodied reasoning), and Gemini Robotics On-Device 2. The ER 2 model acts as a high-level planner, processing live video feeds from robot cameras to track progress and coordinate multi-robot tasks. The on-device model is optimized for local execution and rapid adaptation. The system has demonstrated whole-body control and five-fingered dexterity on the Apptronik Apollo 2 humanoid.
- • Gemini Robotics ER 2 is an upgraded embodied reasoning model based on Gemini 3.5 Flash with a 128k context window.
- • The ER 2 model is integrated with the Gemini Live API and is available to developers starting today.
- • Gemini Robotics On-Device 2 is built on Gemma technology and can adapt to new robot embodiments in a few hours with fewer than 200 examples.
- • The release includes the ASIMOV-Agentic safety benchmark on Hugging Face under a CC-BY-4.0 license.
Developers can now build physical agent workflows using live video feeds and multi-step reasoning via the Gemini Live API.
2. Researchers Distill DeepSeek V4 Flash into GPT-OSS-120B
Using an evolved distillation method based on HINT-SD, researchers successfully transferred complex financial reasoning capabilities to open-source models. The resulting 120B model completes 98.7% of problems within an 8k token budget on a single H100 GPU, costing approximately $0.00026 per query. The team also released the LineageEval framework on GitHub to help developers study censorship transfer in distilled models.
- • Researchers distilled finance reasoning from DeepSeek V4 Flash into a GPT-OSS-120B model.
- • The 120B model scored 83.61% on the FinanceReasoning benchmark, outperforming Kimi K3.
- • Open weights were released for a 20B finance model that runs on a single 80GB GPU.
- • A study of 152 matched prompt pairs showed that teacher model censorship did not transfer to the distilled students.
Developers can access highly specialized, low-cost finance models and utilize the new LineageEval framework to audit censorship transfer.
3. Inkling-Small 276B MoE Model Released with 1M Context
Inkling-Small offers a high-capacity Mixture-of-Experts architecture with a massive 1M token context window. By activating only 12 billion parameters per token, the model balances reasoning depth with computational efficiency. Developers can experiment with local deployment of this model using specialized llama.cpp builds that support CPU offloading for GGUF quantizations.
- • Inkling-Small features 276 billion total parameters and 12 billion active parameters.
- • The model supports a 1 million token context window.
- • It is available in NVFP4 and GGUF formats on Hugging Face.
- • A developmental branch of llama.cpp supports running Unsloth's GGUF quantization on CUDA with CPU offloading.
Developers can run a long-context MoE model locally using a developmental branch of llama.cpp with CUDA and CPU offloading.
4. LG AI Research Releases K-EXAONE 2.0 750B Model
Developed under Phase 2 of Korea's Sovereign AI Foundation Model Project, K-EXAONE 2.0 is three times larger than its predecessor. The model features robust safety alignment, scoring 94.6 on safety benchmarks, and demonstrates competitive performance in coding and agentic tool use. Its Apache 2.0 license makes it highly accessible, though its massive size presents significant hosting requirements.
- • K-EXAONE 2.0 is a 750-billion-parameter model released under the Apache 2.0 license.
- • The model supports 10 languages and scored 94.4 on the OpenAI-MRCR long-context benchmark.
- • It achieved a score of 14.2 on the Tau3-Bench Banking benchmark for agentic tool use, outperforming Qwen 3.5.
- • Coding performance shows an average 30% increase across core metrics compared to version 1.
The release provides a massive, open-weights foundation model with strong agentic tool-use and long-context capabilities.
5. Baseten Adds Vision Capabilities to GLM 5.2 via New Merged Model
Following the recent performance optimization of the GLM 5.2 API, Baseten has expanded the model's utility by releasing GLM-5.2-Vision-NVFP4. This merged model integrates the vision encoder from Kimi k2.6 into the GLM 5.2 architecture, enabling multimodal processing. The release is available on OpenRouter, addressing previous developer feedback regarding the lack of native vision support in the original GLM 5.2 release.
- • GLM-5.2-Vision-NVFP4 is now available for inference on OpenRouter.
- • The model is a merge of the GLM 5.2 architecture and the Kimi k2.6 vision encoder.
- • This release adds multimodal capabilities to the GLM 5.2 family, addressing previous limitations.
Developers can now utilize a vision-enabled version of GLM 5.2 for multimodal applications, building on the existing GLM 5.2 infrastructure.
6. SpaceXAI Launches Grok Voice Think Fast 2.0
SpaceXAI has launched Grok Voice Think Fast 2.0, an updated voice model designed to increase reliability for production customer workflows. The model is available via the Agent Builder at $0.09 per audio minute. Developers currently integrating the `grok-voice-latest` model should prepare for the automatic transition to this new version on August 5.
- • Grok Voice Think Fast 2.0 is now available on SpaceXAI's Agent Builder.
- • The service is priced at $0.09 per audio minute.
- • The grok-voice-latest API endpoint will automatically transition to the new version on August 5.
- • The release focuses on improving dependability in real-world customer workflows.
Developers using Grok Voice can expect improved reliability in production workflows, with a scheduled API transition coming next week.
7. Token Saver MCP Extension Cuts Claude PDF Token Costs by 90-99%
Token Saver provides a zero-configuration, secure way to query local PDF documents using Claude 3.5 Sonnet and Claude 3 Opus. Operating through an 8-stage processing pipeline, the extension runs entirely locally via standard I/O, ensuring zero file uploads to external servers. It features user-configured folder allowlisting and network isolation, making it highly suitable for enterprise environments.
- • Token Saver is an open-source Model Context Protocol (MCP) extension for Claude Desktop (v1.0).
- • The tool implements a local hybrid RAG system using BM25 keyword matching (0.4) and all-MiniLM-L6-v2 semantic search (0.6).
- • It reduces LLM token consumption by 92% to 99% by extracting and sending only relevant document passages.
- • The extension is distributed as a .mcpb bundle, requiring no Python environment or terminal configuration.
Developers can drastically reduce Claude API costs when working with large local PDFs by sending only highly relevant passages to the model.
8. Agent-Manager TUI Simplifies Multi-Agent Execution in Tmux
Agent-manager offers a powerful terminal-based control plane for developers running multiple AI coding agents. By isolating each agent in its own tmux session, the tool allows developers to monitor real-time system metrics, group agents into project trees, and send review comments back to agents as single prompts. It also features automatic status detection via pane polling or hook events for Claude Code.
- • Agent-manager runs coding agents like Claude Code, Codex, and OpenCode in individual tmux sessions.
- • The tool provides a unified TUI to view live status, group agents, and send prompts without attaching to sessions.
- • It supports full-file diff reviews with syntax highlighting and includes an MCP server for agent-driven session management.
- • Installation is supported via Homebrew, shell script, AUR, mise, or Go.
Developers can orchestrate, prompt, and review diffs across multiple concurrent coding agents from a single unified terminal interface.
9. Deep Agents v0.7 Cuts Base Input Tokens by 65%
The release of Deep Agents v0.7 focuses on optimizing runtime efficiency for agentic workflows. By restructuring how base inputs are handled, the framework cuts input token usage by 65%. This optimization allows developers to run complex agent loops at a fraction of the previous API cost while maintaining identical output quality.
- • Deep Agents v0.7 reduces base input tokens by 65% compared to previous versions.
- • The token reduction is achieved with no degradation in agent performance.
- • The update is designed to directly improve cost and token efficiency for developers.
Developers using Deep Agents can immediately lower their API costs and improve token efficiency without sacrificing agent performance.
10. Perplexity Open-Sources Numbat Agent Security Suite
Perplexity's Numbat addresses the unique security challenges of running autonomous AI agents on client devices. By integrating directly with agent execution harnesses, the open-source suite provides developers with tools to monitor agent behavior, prevent accidental meltdowns, and mitigate security risks at the endpoint level.
- • Numbat is an open-source security suite developed by Perplexity specifically for AI agents.
- • The suite targets security risks associated with deploying autonomous agents on client endpoints.
- • Numbat integrates directly with agent harnesses to detect and mitigate runtime security incidents.
Developers can integrate Numbat into their agent harnesses to prevent, detect, and mitigate accidental agent meltdowns at the client level.
11. Hush Security Launches Identity Gateway for Autonomous AI Agents
Israeli startup Hush Security is tackling the security challenges of the rapidly growing AI agent ecosystem. Because autonomous agents often inherit broad human credentials or use long-lived API keys, they present significant security risks. Hush Security's Identity Gateway platform allows organizations to assign unique identities to agents and enforce granular, task-specific permissions. A free tier is available to help developers gain runtime visibility and implement access controls.
- • Hush Security raised $30 million in Series A funding to expand its platform to govern autonomous AI agents.
- • The Identity Gateway platform discovers AI agents and enforces 'least agency' by brokering task-specific permissions.
- • A free plan is available, providing runtime visibility, risk analysis, and identity-based access controls.
- • The platform addresses security risks from agents operating on inherited human credentials or long-lived API keys.
Developers can secure agentic deployments by assigning unique non-human identities and enforcing task-specific, brokered permissions.
12. Open-Source Agent Skill Enforces Simplified Technical English
This open-source agent skill provides specific adaptations for error messages, runbooks, incident reports, and release notes. By enforcing Simplified Technical English, the skill not only improves documentation clarity but also consistently decreases output token counts across all tested models. Developers can easily add the skill to their existing agent configurations or system prompts.
- • The open-source skill forces LLMs to write in ASD-STE100 Simplified Technical English (STE).
- • It is compatible with Claude Code, Cursor, VS Code Copilot, and other major agent harnesses.
- • Testing showed a 72.9% reduction in STE violations and reduced mean sentence length from 11.2 to 9.7 words.
- • The skill is licensed under the MIT license and can be installed via `npx skills add AminBlg/SimpleEnglish`.
Developers can integrate this skill into Claude Code, Cursor, or Copilot to automatically generate clearer, more concise documentation and save tokens.
13. Tencent Open-Sources AngelSpec Speculative Decoding Framework
Tencent's AngelSpec provides a unified framework for training speculative-decoding draft models. It offers two complementary architectures: an MTP model for conversational data and a block-diffusion model (DFly) optimized for code and mathematics. AngelSpec includes advanced features like Ulysses sequence parallelism for long-context training and document-aware sequence packing, allowing developers to deploy highly optimized local inference pipelines.
- • AngelSpec is an open-source, torch-native framework built on TorchSpec for training speculative-decoding draft models.
- • The framework supports autoregressive multi-token prediction (MTP) and the block-parallel DFlash family.
- • On the HY3-295B-A21B model, the DFly-8 draft model delivers a 1.98x to 2.40x speedup over standard autoregressive decoding.
- • AngelSpec integrates with vLLM, SGLang, and Hugging Face Transformers, and includes seven pre-trained checkpoints.
Developers can train custom draft models to achieve up to a 2.4x speedup in local LLM inference without modifying their target models.
14. Cisco Launches Free AI Supply Chain Provenance Explorer
Cisco's AI Supply Chain Provenance Explorer addresses the lack of verification in open-source AI, where up to 69% of open models have unverified lineage. By analyzing weight-level signals like Embedding Anchor Similarity, the tool bypasses self-reported metadata to establish true model origins. This verification is increasingly critical as the European Commission begins enforcing the EU AI Act on August 2, 2026, which carries steep fines for unverified modifications of open models.
- • The AI Supply Chain Provenance Explorer is a free public database covering nearly 900 open-source models.
- • The tool uses static fingerprinting and behavioral-similarity analysis to verify model lineage with 96.4% accuracy.
- • It provides searchable data on provider headquarters, license restrictions, and scanned files.
- • The database is available publicly at provenance.aidefense.cisco.com without requiring a Cisco account.
Developers can verify the true lineage and licensing of open-source models to ensure compliance with upcoming regulations like the EU AI Act.
15. Escha-W2 Quantized Qwen MoE Model Released for Local GPUs
Escha-W2 makes hosting a massive 256-expert Mixture-of-Experts model accessible on consumer-grade hardware. By compressing the Qwen3.6-35B-A3B model down to a 2-bit quantization, the build fits within 12.3 GB of disk space. It comes pre-packaged with the necessary serving infrastructure to expose an OpenAI-compatible HTTP API out of the box.
- • Escha-W2 is a 2-bit quantized build of the Qwen3.6-35B-A3B MoE model, which features 256 experts.
- • The model is packaged with tools to serve it locally via an OpenAI-compatible HTTP API.
- • It occupies 12.3 GB of disk space and runs on a single consumer GPU with 16 GB or 24 GB of VRAM.
Developers can host a 256-expert Mixture-of-Experts model locally on 16GB or 24GB VRAM hardware using an OpenAI-compatible API.
16. MindControl Benchmarks Show 50% Token Reduction in Coding Tasks
Following the initial release of the MindControl llama.cpp fork, developer Laurence Hardman has published benchmark results quantifying its performance. By replacing hard-truncation with sampler-level guided reasoning budgets, the tool allows models like Qwen3.6-27B to maintain high accuracy on coding benchmarks while reducing token usage by over 50% on LiveCodeBench.
- • MindControl benchmarks on Qwen3.6-27B show a 50% reduction in token consumption on LiveCodeBench.
- • The most constrained configuration achieved a 95.7% score on HumanEval+.
- • The technique maintains high accuracy while optimizing token budgets, though performance costs were noted on the hardest problem subsets.
These results confirm that the previously released MindControl fork can significantly improve the efficiency of local model inference for coding tasks without aggregate accuracy penalties.
17. OpenAI Adjusts GPT-5.6 Pricing and Adds High-Throughput 'Fast Mode'
Building on the general availability of the GPT-5.6 model family released earlier this month, OpenAI has implemented significant pricing adjustments to remain competitive. The company has slashed API pricing for the cost-optimized Luna model by 80% and the balanced Terra model by 20%. Additionally, OpenAI has introduced a new 'Fast mode' for the flagship Sol model, providing 2.5x higher throughput for latency-sensitive production workflows at a premium price point.
- • GPT-5.6 Luna prices reduced by 80% to $1.40/M tokens combined.
- • GPT-5.6 Terra prices reduced by 20% to $14.00/M tokens combined.
- • New 'Fast mode' for GPT-5.6 Sol offers 2.5x throughput at $70/M tokens combined.
- • Standard pricing for GPT-5.6 Sol remains unchanged.
These updates provide developers with more flexible cost and performance options for the GPT-5.6 series, allowing for cheaper lightweight reasoning or higher-throughput production capabilities.