Records found: 6
'A practical comparison of the top 6 LLM inference runtimes in 2025, highlighting design tradeoffs, KV cache strategies, performance profiles, and where each engine fits in production.'
'MiniMax-M2 is a compact MoE released with open weights on Hugging Face, optimized for coding and agentic tool use with about 10B active parameters per token and claims of ~2x speed at ~8% of Claude Sonnet cost.'
kvcached provides a virtualized, elastic KV cache for LLM serving on shared GPUs, reducing memory waste and speeding activation across colocated models.
'GLM-4.6 expands context to 200K tokens, reduces token consumption on real-world coding benchmarks, and offers open weights for local inference and research.'
'BentoML launched llm-optimizer to automate benchmarking and tuning of self-hosted LLMs and published a browser-based LLM Performance Explorer with pre-computed results.'
Researchers at UC Berkeley and UCSF have developed Adaptive Parallel Reasoning, a novel method that allows large language models to dynamically distribute inference tasks across parallel threads, enhancing reasoning performance without exceeding context window limits.