Running DeepSeek V4-Flash with 1M Context on 2x RTX PRO 6000: Concurrency Up 1.5x
dee_hw · x · 2026-08-02
A developer corrected the previous assumption that a 2× RTX PRO 6000 rig is capped at a 131K context length, confirming it can run the full 1M context window for DeepSeek V4-Flash.
- VRAM Optimization: Thanks to sliding-window attention, the 1M context isn't a physical allocation. The actual KV cache footprint is only 16K tokens.
- Performance: Successfully retrieved a needle at 543K depth on native FP4.
- Trade-offs: Lowering numbatchedtokens increases concurrency by 1.52x, but prefill time jumps significantly from 11.9s to 120.3s.
Related event: DeepSeek V4 Flash Sparks a Wave of Local Deployment Tests(26 posts)→
More from Infra
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16
- Google Raises 2026 Capital Spending to Up to $205 Billion — emmanuelvivier · 2026-10-16
- Google open-sources AX, a Kubernetes-style orchestrator for billions of AI agent tasks — thisdudelikesAI · 2026-10-16