AgentX benchmark: lossless-CSF hits 11.4M KV cache tokens, 8.7x DeepSeek V4.1 Flash baseline
TheZachMueller · x · 2026-10-30
A local-inference comparison run by @keennay: on 4× RTX PRO 6000 Blackwell GPUs, the SemiAnalysis AgentX benchmark (Claude Code 256k trace dataset) pits DeepSeek-V4.1-Flash-1017 DSpark 7 against Local Inference Lab's lossless-CSF variant (vLLM, DSpark + n-gram + RAM offload). CSF's checkpoint has far more headroom — 8.7x the total KV cache tokens at 11.4M vs 1.3M — with clear throughput gains at high concurrency (C16/C32) and near-parity at low concurrency; one empty-response error each at C8/C32. Charts cover throughput-vs-interactivity, TTFT/ITL latency and effective concurrency curves.
More from Infra
- NVIDIA's Nemotron 3.5 Super hits 51.78% on SWE-bench Pro with 4.6x fewer active params — dair_ai · 2026-10-30
- Bonsai Compresses Qwen 27B from 55.6GB to 5.9GB, Retains 98.2% Benchmark Performance — nicolechirps · 2026-10-30
- Weaviate: Use Slow-Query Log and Profiling to Fix Slow Hybrid Search Queries — CShorten30 · 2026-10-30
- Nearly $1 Trillion Wiped From US Stocks at Open as AI Chip Stocks Lead Selloff — CurieuxExplorer · 2026-10-30
- Anthropic eyes $2T valuation on $100B raise, with $431B compute commitments and circular vendor financing — LexSokolin · 2026-10-30
- Cloudflare private origins goes GA: WAF and caching for internal APIs and MCP servers — Cloudflare Blog · 2026-10-30