Benchmark
ESP32-P4 vs ESP32-S3: Energy per Token, Measured at the USB Port
Board-level energy per token for a TinyStories model: ESP32-P4 7.9 mJ, ESP32-S3 12.4-14.7 mJ. The P4 draws more power and still uses about 46% less per token.
LLM on a Microcontroller: ESP32-S3 at 22.6 tok/s
esp32-ai's 28.9M-parameter model on an $8 ESP32-S3: 10.3 to 22.6 tok/s from five exact patches, same output to the byte. PSRAM bandwidth is the limit.
Strata vs llama.cpp on an RTX 3090: 78 tok/s vs 18 on the Full Flash-Next
The full Qwen3.8-Flash-Next on a 24 GB RTX 3090 with 64 GB of DDR4. Strata decodes at 78 tok/s against llama.cpp's 18, and reads prompts about 3x faster.
Does RAM Speed Matter for Local LLMs? 2133 vs 3000, Measured
DDR4-2133 vs 3000 on an RTX 3090 box: no change for models that fit on the GPU, 13 to 29% faster with MoE experts offloaded. Plus why 3200 wouldn't train.
Qwen 3.8 35B-A3B Is a Distill: Tested Against Qwen 3.6
There is no official Qwen 3.8 35B-A3B. I ran empero-ai's distill against Qwen 3.6 35B on an RTX 3090: same score, same speed, a fifth of the reasoning tokens.
Strata vs llama.cpp on a 3060: The 7x Is Real. Two Flags Halve It
Qwen3.8-Flash-Next Coder on a 12 GB RTX 3060 and 32 GB of DDR4. Strata prefills 6-7x faster than stock llama.cpp. One batch setting closes most of that.
Bonsai 2 vs Qwen3.8-27B on an RTX 3090: Speed Doubled, Scope Broke
Bonsai 2 27B, PrismML's ternary Qwen3.8, on a 3090: 1.8x the decode of Q4 from a 7.2 GB file, then 12 of 47 against 19 on a routing split. One field broke.
Jev Mode on a 3090: 26 ms per Token You Don't Write
Jev-style parallel decisions in llama.cpp on Qwen3.6-27B, RTX 3090: 1.3x faster at four tokens, 5.5x at fifty, 21 vs 23 of 47. The gain is the tokens you skip.
GTX 1650 vs RTX 3060 on a 35B MoE: What the Card Buys
A $60-class GTX 1650 4 GB runs Qwen3.6-35B-A3B at 20 tok/s on 32 GB of RAM. The RTX 3060 in the same slot does 28 at the same setting and 39 tuned. Measured.
A 177B Model on a 3060: The 32 GB Number Nobody Measured
Qwen3.8-Flash-Next on an RTX 3090 and a 3060. Stock, the 3090 matches the video's 3060. One flag buys 34 to 40 percent. A real 32 GB box: 8.7 tok/s, not 22.
Nvidia PAIR Bought Me 7 Percent. My Own Router Cost Me Half.
Nvidia's new home-network AI router split 20 requests across an RTX 3090 and a 3060 for 1.07x over the 3090, then dropped half a 27B queue. Mine: 0.52x.
LoRA Skill Compilation Is a Double-Headed Coin Flip
Ten LoRA seeds on identical data spread 3.62 points, against 4.65 for prompt compilation. Not one beat the no-adapter base. Measured on an RTX 3090.
I Got the AI Result I Wanted. Then I Ran It Nine More Times
A frontier model read my logs and wrote a skill that beat my local 27B baseline by 10.6 points. Nine more compilation runs showed the number was fake.
Ornith 1.5 35B vs Qwen 3.6 on RTX 3090: Speed Tested
Firsthand A-B-B-A bench of Ornith 1.5-35B-A3B against Qwen 3.6-35B-A3B on one RTX 3090. Generation, prefill, VRAM, and the noise floor under all three.
Why Qwen 3.8 27B Feels Slow: Reasoning Tokens Measured
Qwen 3.8 27B generates at full speed on a 3090 and still crawls. Four runs, two models, two seeds: 92.8% of output is thinking, and the spread runs 320x to 542x.
Qwen 3.8 27B vs 3.6 on RTX 3090: Speed and Quality Tested
Firsthand benchmarks of Qwen 3.8-27B against 3.6-27B on one RTX 3090: generation within a percent, VRAM +254 MiB, and HumanEval pass@1 a statistical tie.
Best 24GB Backend Shootout: ik_llama vs BeeLlama vs llama.cpp
ik_llama and BeeLlama both finish in 22-23s on the am17an 9-prompt harness vs mainline llama.cpp's 37s — 1.66x and 1.62x speedups via opposite strategies.
Wicked Fast Qwen 3.6 27B: 60 tok/s with MTP on RTX 3090 (2026)
Firsthand bench: 60 tok/s on Qwen 3.6 27B Q4_K_M with MTP on a single RTX 3090 — 1.86x wall-clock speedup over baseline. PR #22673 progress May 6 → May 19.
Wicked Fast Gemma 4 vs Qwen 3.6 on RTX 3090: 3.10x Tested
Same RTX 3090, same llama.cpp build, same bench. Gemma 4 26B-A4B Q4_K_XL: 128 tok/s mean. Qwen 3.6-27B Q4_K_M: 41 tok/s. 3.10x faster, firsthand.
DFlash vs MTP on RTX 3090: I Tested Both Locally
Firsthand head-to-head bench of DFlash + DDTree against MTP (PR #22673) on a single RTX 3090, same Qwen 3.6-27B target. Real numbers, both backends.
Best Way to Run Qwen 3.5 on Mac: MLX vs Ollama Speed Test
MLX runs Qwen 3.5 up to 2x faster than Ollama on Apple Silicon. Head-to-head benchmarks on M1 through M4, with setup instructions for both.
A weekly email with every new guide and measured benchmark.
Subscribe — free, no spam