Building MapSleuth, a reverse engineering tool that finds and decodes the calibration maps hidden in car ECU firmware, running on local AI on 2 DGX Sparks.
Just moved its model from vLLM to TensorFold: 57 vs ~40 tok/s, first token 0.18s vs ~1.25s. 🕵️🔧
#ECUtuning#LocalAI
@sfxnz 100%. Same 2x Spark, same model, just a better engine: 57 vs ~40 tok/s and first token 0.18s vs ~1.25s for me. The community basically handed my ECU reverse engineering project, MapSleuth, a free hardware upgrade.
@niklaslenz_ai I’m at around 60% coverage on MED17 and MG1 ECUs right now. Still working on improvements. The reference files I’m using to check against so far are Audi RS6 (4.0TT), Porsche Panamera (3.0T), and Lamborghini Huracan/Audi R8 (5.2 V10)
@hardwear_io Love seeing some car hacking stuff on X. I'm on the calibration side: building MapSleuth, a reverse engineering tool for car ECU maps, running fully local on 2x DGX Spark. Anyone there digging into ECU firmware this year?
@plotarmordev@ashxhart Big release. On 2x Spark with Flash-Next I'm already at 130 vs 102 tok/s at 4 concurrent vs vLLM, so shared-prefix gains at 8 streams are exactly what agent workloads need. The read-ahead + fail-fast on short files is a nice touch too.
@MiaAI_lab Thanks Mia! Your Flash-Next recipe is what runs MapSleuth (my reverse engineering tool for car ECU maps) on 2x DGX Spark. Moving to TensorFold took single-request speed from ~40 to 57 tok/s for us. Keen to try the lower TTFT.
@ashxhart Congrats on 1.2.10! Multi-stream gains on GB10 are great news for Spark owners. On 2x Spark I already saw 130 vs 102 tok/s at 4 concurrent vs vLLM, so I'm curious where 8 streams lands now. Will update and report back.
@jasona 2x DGX Spark, TensorFold serving Qwen3.8-Flash-Next, Hermes Agent on top. Been solid. The big win for agent loops was first token: 0.18s vs ~1.25s avg on vLLM, so tool-call turns feel snappy.
@ivanfioravanti 100% agree. Same 2x DGX Spark, same Qwen3.8-Flash-Next, only swapping vLLM for TensorFold: 57 vs ~40 tok/s single request, first token 0.18s vs ~1.25s. The recipe really is half the box.
@squarrier Love the shift from single-stream tok/s to long conversations. That's where TensorFold surprised me too: on 2x DGX Spark with Qwen3.8-Flash-Next I measured ~30% less waiting than vLLM at ~170k context. How many long convos can you hold?
@ashxhart Same rabbit hole. I'm building MapSleuth, a reverse engineering tool for car ECU maps, running fully local on 2x DGX Spark with Qwen3.8-Flash-Next on TensorFold. Verifying what the model finds is the hard part, and it's where most of my effort goes. Thanks for TensorFold!
Building MapSleuth, a reverse engineering tool that finds and decodes the calibration maps hidden in car ECU firmware, running on local AI on 2 DGX Sparks.
Just moved its model from vLLM to TensorFold: 57 vs ~40 tok/s, first token 0.18s vs ~1.25s. 🕵️🔧
#ECUtuning#LocalAI