Agentic AI News Aug. 2, 2026

Intelligence on par with March 2026. DeepSeek V4 Flash (0731) achieves an intelligence index score of 50, nearly matching the 51 of frontier models from five months ago. On consumer hardware, it runs at usable speeds: 44 t/s on an RTX PRO 6000 eGPU, 15 t/s on triple MI50s, and even 4.5 t/s on dual RTX 3060s with RAM offloading, thanks to clever quantization and partial offloading.

• DeepSeek-V4-Flash-0731: Models you can run locally now have the intelligence score of the top frontier model from March 2026(opens in a new tab)
• DeepSeek-V4-Flash-0731 on Bosgame M5 with RTX PRO 6000 Max-Q eGPU(opens in a new tab)
• Ran DS V4-Flash-0731 Locally on 3xMI50 32GB @ ~15 t/s TG(opens in a new tab)
• DeepSeek V4 Flash 0731 IQ2_M benchmark for Dual 3060 and 96GB RAM ≈ 3.5 tok/s.(opens in a new tab)

Tool calling fixed, rule following still flaky. After community reports of broken tool calls, a llama.cpp fix was merged, and workarounds for Unsloth Studio were shared. However, users find that the model often ignores system prompts and skills, storing them as compressed summaries rather than adhering strictly, which undermines its reliability for agent tasks that require precise instruction following.

• Fix for Deep Seek v4 Flash 0731 tool calling has been added to llama cpp(opens in a new tab)
• DeepSeek V4 Flash 0731 local setup gotcha: model, tool call & config setting(opens in a new tab)
• Deepseek v4 flash 0731 still not holding up.(opens in a new tab)

Community benchmarks for Hermes agents. LocalLLaMA users discuss the best benchmarks for evaluating agent-capable models like Hermes, with many feeling that established leaderboards like SWE-bench don't apply to their home-assistant or document-processing workflows.

• What’s the community’s favorite benchmark to validate performance?(opens in a new tab)

Coding-only leaderboards criticized. A call for more diverse benchmarks argues that recent leaderboards disproportionately focus on coding tasks, leaving language learning, creative writing, and STEM reasoning underrepresented, making it hard to gauge model performance for non-coding agent use cases.

• Why are almost all new benchmarks and leaderboards coding focused?(opens in a new tab)

DIY domain-specific benchmark kit. A Redditor built a collection of over 30 small, domain-specific tests (e.g., common sense, car wash simulation) and a web tool to compare model responses, aiming to fill the gap between broad benchmarks and personal use cases.

• A collection of small domain-specific benchmarks for local models (30+ and growing)(opens in a new tab)

Astra cracks decade-old math problems. OpenAI’s next model family, Astra, solved ten previously open problems in math and theoretical computer science, with formal Lean proofs and a cost of under $2,000 per solution. The results span geometry, coding theory, and quantum complexity, and have already inspired further human research.

• Ten advances in mathematics and theoretical computer science(opens in a new tab)
• OpenAI announces its "next major model" Astra by dropping ten previously unsolved math solutions(opens in a new tab)

Coding agents revive aging research software. A field study with OpenAI and academic partners shows that AI coding agents like GPT-5.5 and Claude Code can modernize fragile scientific software, achieving over 60x speedups in some cases, but the work shifts the burden to verifying correctness—the agents can’t judge the science.

• AI coding agents can modernize research software but can't judge if the science is right(opens in a new tab)

Self-spreading prompt injection worm attacks Copilot. A security researcher demonstrated a worm that uses hidden white-on-white text in Word documents to inject malicious instructions into Microsoft Copilot, which then copies the payload into new files. Microsoft confirmed the behavior but hasn't fixed it after 144 days.

• A security researcher built a self-spreading worm that hides inside Word docs and hijacks Microsoft Copilot(opens in a new tab)

EU AI Act labeling mandate begins. From August 2, 2026, the EU requires that AI-generated images, video, audio, and text be clearly labeled. The regulation may impact agents that autonomously publish content or create media.

• EU AI Act takes effect tomorrow, August 2, 2026. 🤡(opens in a new tab)

Minnesota’s ‘nudify’ ban survives xAI legal challenge. A federal judge refused xAI’s attempt to temporarily block Minnesota’s law against apps that generate non-consensual explicit images, noting xAI waited three months to sue. The law takes effect while the lawsuit continues.

• Judge denies xAI’s request to block Minnesota ban on ‘nudify’ apps(opens in a new tab)

Kimi K3 runs on a single CPU, slowly. Two developers independently wrote minimal C inference engines to run the 1.56 TB MoE model Kimi K3 on disk with no GPUs. One achieved a token every 30 seconds using 8 GB RAM and NVMe streaming; another got 0.5 t/s with 29 GB RAM, all experts read on demand.

• I pushed Kimi K3 onto one CPU with 8 GB of RAM(opens in a new tab)
• Weight-Aware Streaming Tensor Engine: run Kimi K3 using 29 GB of RAM at 0.50 tok/s(opens in a new tab)

LongCat-Flash-Lite-Sparse now available. This new release replaces dense attention with LongCat Sparse Attention (LSA) and supports up to 1 million tokens of context natively, making it suitable for long-document agent tasks.

• LongCat-Flash-Lite-Sparse Is Now Available for Download(opens in a new tab)

Laguna S 2.1 updates context and quantization. Poolside released new FP8 and NVFP4 checkpoints for Laguna S 2.1, bumping the default context size to 1M tokens and hoping to fix a looping issue that had hampered development workflows.

• New official weights for Laguna S 2.1 FP8 & NVFP4 are now available(opens in a new tab)