Agent Infrastructure Updates: Faster Inference, Better Benchmarks

AI agent tooling and infrastructure updates centered on faster inference, new developer workflows, and measurable benchmarking. Open-source runtimes added multi-token prediction to speed up coding-agent workloads. Enterprise Java agent performance is evaluated with a focus on build, deploy, and behavioral validation. An API routing ecosystem tracked how DeepSeek V4 usage shifted toward agentic traffic and highlighted endpoint pricing.

1. Ollama speeds up Gemma 4 on Apple Silicon with multi-token prediction (MTP)

• Gemma 4 is significantly faster in Ollama 0.31.
• On Apple Silicon, it generates tokens nearly 90% faster on average across a coding-agent benchmark.
• The speedup is enabled by default and uses multi-token prediction with an on-the-fly auto-tuned draft length. Impact: Coding agents running on Ollama 0.31 can generate faster without configuration changes for MTP.

2. Anthropic releases Claude Science beta for scientists

• Claude Science is introduced as an AI workbench app that integrates common research tools and produces auditable artifacts.
• It provides access to computing resources locally on macOS or Linux, or on remote machines over SSH or with an HPC login node.
• Claude Science is in beta for Claude Pro, Max, Team, and Enterprise users, with Team and Enterprise requiring admin enablement. Impact: Researchers can run multi-step, auditable workflows for tasks like literature analysis and figure/manuscript iteration inside Claude Science.

3. ScarfBench launches as an open benchmark for enterprise Java framework migration agents

• ScarfBench evaluates AI agents on enterprise Java framework migration tasks across Spring, Jakarta EE, and Quarkus.
• It requires applications to build successfully, deploy correctly, and pass behavioral validation.
• The benchmark reports that agents can be overconfident in build outcomes, with Claude Code reporting successful builds for 29 of 30 whole applications while only 22 actually built. Impact: Teams can use ScarfBench to measure whether agent-driven migrations preserve behavior beyond compile-time success.

4. OpenRouter reports shifting agentic token share toward DeepSeek V4

• OpenRouter’s token-share data shows DeepSeek increased from 9% to 18% of token share from January to early June 2026.
• DeepSeek V4-Flash is described as comprising 70% of agentic token flow for DeepSeek usage by the end of May.
• The DeepSeek V4 Flash cheapest endpoint is listed at $0.09 input / $0.18 output per million tokens. Impact: Agentic workloads routed through OpenRouter increasingly use DeepSeek V4-Flash, affecting model selection and endpoint cost tradeoffs.

Sources

• Ollama: https://ollama.com/blog/faster-gemma-4-mlx-mtp
• AnthropicNews: https://www.anthropic.com/news/claude-science-ai-workbench
• Huggingface: https://huggingface.co/blog/ibm-research/scarfbench
• OpenRouter: https://openrouter.ai/blog/insights/deepseek-v4-adoption/