Hermes Agent Hits #1, llama.cpp Gets RDNA3 Tensor Cores, and MiniCPM-V Runs on Your Phone

News

Nous Research's Hermes Agent Dethrones OpenClaw as the Most-Used Open-Source AI Agent on OpenRouter

Three months after launch, Hermes Agent hit 224 billion daily tokens on OpenRouter, beating OpenClaw's 186 billion and claiming the global #1 spot. The v0.13.0 'Tenacity' release shipped May 7 with a durable multi-agent Kanban board, zombie-worker recovery, and 8 P0 security fixes -- this is what an open-source agent runtime looks like when it ships like it means it.

DeepSeek V4 Launches Open-Weight with 1M Token Context and MIT License at $0.30/M Output Tokens

DeepSeek V4 Pro (1.6T params, 49B active) and V4 Flash (284B total, 13B active) landed April 24 with open weights on HuggingFace under MIT, a 1M-token default context, and API pricing that comes in at a fraction of GPT-5.5. The hybrid attention architecture (CSA + HCA) cuts KV cache requirements by 90% versus V3.2 at 1M context -- the most consequential open-weight release for long-context agentic workflows this quarter.

Tools

llama.cpp Adds RDNA3 Tensor Core Support to Flash Attention Kernel

The latest llama.cpp release, pushed 16 hours ago, adds RDNA3 tensor core support to the CUDA mma FA kernel, with tuned parameters for RDNA3, RDNA4, and CDNA1 architectures. This is a meaningful win for AMD GPU users running local inference -- RDNA3 cards like the RX 7900 XTX have had patchy attention kernel performance and this closes the gap.

OpenBMB Releases MiniCPM-V 4.6: A 1.3B Multimodal Model That Runs on iOS, Android, and HarmonyOS

MiniCPM-V 4.6 dropped May 11 as a 1.3B-parameter vision-language model with image and video understanding, 262K token context, and full on-device deployment across iOS, Android, and HarmonyOS under Apache 2.0. Artificial Analysis confirmed it's the highest-scoring open-weight model under 2B parameters on their Intelligence Index, and it's "the most edge-deployment-friendly model to date" per the team's own release notes.

Analysis

Kimi K2.6 Holds the Open-Weight Crown: 1T-Param MoE with 300-Agent Swarms and 13-Hour Coding Runs

Moonshot AI's Kimi K2.6 (1T total params, 32B active) scored 58.6% on SWE-Bench Pro, edging GPT-5.4, and landed #4 on the Artificial Analysis Intelligence Index behind only Anthropic, Google, and OpenAI. The headline feature is Agent Swarm scaling to 300 sub-agents with 4,000 coordinated steps and documented 12-hour autonomous coding sessions -- open weights under modified MIT on HuggingFace.

DeepSeek V4 Resets the Price Floor for Frontier-Grade Coding Intelligence

DeepSeek V4 scores 80.6% on SWE-bench Verified at $0.30/M output tokens -- "83x cheaper than Claude Opus 4.7" while sitting at near-frontier performance. The practical implication for builders: the cost constraint on agentic coding pipelines just moved an order of magnitude, and the remaining gap vs. closed models is primarily in long-context retrieval and multi-constraint instruction following.