Why Is Strata So Fast? And What's the Catch? Strata Tested!

Original author：Weights & Wonders

Original source：https://www.youtube.com/embed/8kzSGQzGigc

A friend asked me to host Qwen 3.8 Flash Next for him: a 125-billion-parameter model, on a machine with one RTX 4070 (12 GB). So I tried Strata, the free engine people are calling "6× faster than llama.cpp". It works. Writing ran at about 35 tokens a second against llama.cpp's 15, and a 100,000-token codebase was read in 2 min 22 s instead of 24 minutes. Then the logs showed something odd: the graphics card was barely trying. This video explains why. The same reason makes this setup brilliant for one person and no faster than llama.cpp once a second person joins. CHAPTERS 0:00 A 125B model on one 12 GB card 1:16 The request 2:18 How Strata splits the model 4:07 The machine 5:28 One person, one chat 6:39 The long read 8:41 The card that was barely trying 9:56 Two's a crowd 11:05 Verdict THE RESULTS (same model file, same card, both engines) • Short chats: Strata 32.5–38.1 tok/s vs llama.cpp 15.3–16.5 tok/s (stock llama.cpp, no MTP draft) • Reading a 100,000-token C++ codebase: Strata 2 min 22 s vs llama.cpp 24 min (~71 tok/s) • Writing after that 100K read: Strata 47 tok/s (373 of 404 drafted tokens accepted on code) vs llama.cpp 15.5 • 235,000-token document: read in 5 min 5 s; hidden passphrase found. Follow-up question: 11.5 s (229,376 tokens reused from cache, 6,054 new) • Writing as the chat grows: 30.0 tok/s (~1K) → 25.7 (120K) → 24.6 (235K) • GPU power: average 93 W of a 200 W limit (peak 117 W), clocks at full speed (~2,865 MHz) • Where the needed experts ran (single user): 58% already on the card, 4% copied to it, 38% on the CPUs • Two users (parallel on): ~18 tok/s combined vs llama.cpp ~19. Four users: 24.2 vs 23.2 • Two users each pasting ~90,000-token documents at once: 23 min 45 s • Four short chats: 66.5 s in parallel, 49.8 s one after another THE SETUP • GPU: NVIDIA RTX 4070, 12 GB • CPU: 2× Intel Xeon E5-2680 v4 (2016) • RAM: ~170 GB available (the model's experts need ~50 GB; 64 GB total is the practical minimum for this size) • Model: OrcaRouter's Qwen 3.8 Flash Next Uncensored, IQ3_XXS GGUF (85 GB); n-gram table (~29 GB) read from the SSD • Engines: Strata 0.1.40 (262K context, MTP draft from the original Qwen model) vs llama.cpp (stock) • All numbers come from our own logged runs on 7 Oct 2026. WHAT IT COSTS Around 7 Oct 2026, a used RTX 4070 was about $600 and a 64 GB DDR4 kit about $490. RAM prices are moving fast, so check current listings. RAM BY MODEL SIZE (from Strata's docs) 32 GB: Coder version · 48 GB: 2-bit versions · 64 GB: 3-bit versions (used here) · 96 GB+: ~4-bit #LocalAI #Qwen #Strata #LocalLLM #llamacpp #RTX4070 #AI #LLM #NVIDIA #GPU #SelfHosted #HomeLab #OpenSourceAI #MoE #AIServer #strata #strataai