02Local AI
Engines, recipes and a router. All measured.
I contribute to the TensorFold inference engine, publish DGX Spark recipes for DeepSeek, GLM and Qwen, and build LocalRouter. Every number links to its receipt.
- Rate
- 87.9 tok/s
- Tokens
- 0/0
- Elapsed
- 0.00 s
Historical DeepSeek Flash-1017 short decode, 1 stream, 2 Sparks, EXL3 2.8 bpw. This text replays at the published rate. It is not live inference. Receipt ↗
L4 Products
Lockpick.appFind the exposure. Make the repair. Check it again.AutoVisuals.ioSee the car before you buy the parts.L3 Router
LocalRouter~0.9 sMedian voice turn to first audioL2 Recipes
DGX Spark recipes87.9tok/s · DeepSeekDeepSeek, GLM, and Qwen. Measured on real hardware.L1 Engine
TensorFold50Authored commits on the upstream-main snapshotL0 Silicon
NVIDIA DGX Spark 128 GBper unit · GB10 Spark ACX7 · 200 Gb/s Spark B
Packets rise at the selected rate L0 → L4
Projects
What I am building and contributing to right now.
01 / 03 · L1 Engine Major contributor
TensorFold
Native inference. Work that reaches the engine.
I contribute to TensorFold, the Zig inference engine for Apple Silicon and NVIDIA GPUs. My merged work covers CUDA communication, DeepSeek serving and session reliability, memory accounting, and Windows support.
- Built a CUDA communicator interface for NCCL transports, byte gathers, and peer exchanges.
- Improved DeepSeek session handling, CUDA failure handling, and tool-call enforcement.
- Added memory receipts and full token digests to make performance comparisons easier to audit.
Publishedmeasured, not estimated
- Authored commits on the upstream-main snapshot
- 50Receipt ↗
02 / 03 · L2 Recipes Recipe author and performance engineer
DGX Spark recipes
DeepSeek, GLM, and Qwen. Measured on real hardware.
I publish recipes for running DeepSeek, GLM, and Qwen on NVIDIA DGX Spark. Each recipe documents its serving stack, model format, and settings, with measured results and the limits of the setup.
- DeepSeek-V4.1-Flash-1017 on two Sparks with TensorFold. Published speed results describe the previous engine and profile; the current pin is still under qualification.
- GLM-5.3-Flash on two Sparks with TensorFold, with workload-specific benchmark results.
- Qwen3.8 Flash Next on one Spark with exllamav3, TabbyAPI, and MTP speculative decoding.
Published resultstok/s
DeepSeek87.9tok/s
Historical DeepSeek Flash-1017 short decode, 1 stream, 2 Sparks, EXL3 2.8 bpwReceipt ↗
GLM72.4tok/s
GLM-5.3-Flash code decode, 2 Sparks, W20, RigMark mean of 3 run mediansReceipt ↗
Qwen79.9tok/s
Qwen3.8 Flash Next code decode, 1 Spark, greedy, EXL3 3.05 bpw, MTP onReceipt ↗
Conditions differ per row. These are receipts, not a ranking.
03 / 03 · L3 Router Creator and engineer
LocalRouter
Your GPU. Your models. One router.
I built LocalRouter to run chat, voice, image, and video models on a DGX Spark through one Zig daemon. It keeps the voice loop warm, loads other models on demand, and exposes them to agents over MCP and an OpenAI-compatible API.
- Interruptible voice conversations with local speech recognition and synthesis.
- Load models on demand and make room for the next job.
- Early preview, with a beta desktop companion.
Receipts
Every published number on this page, with what was measured and where to check it.
| Value | What was measured | Project | Receipt |
|---|---|---|---|
| 50 | Authored commits on the upstream-main snapshot | TensorFold | Receipt ↗ |
| 87.9tok/s | Historical DeepSeek Flash-1017 short decode, 1 stream, 2 Sparks, EXL3 2.8 bpw | DGX Spark recipes | Receipt ↗ |
| 72.4tok/s | GLM-5.3-Flash code decode, 2 Sparks, W20, RigMark mean of 3 run medians | DGX Spark recipes | Receipt ↗ |
| 79.9tok/s | Qwen3.8 Flash Next code decode, 1 Spark, greedy, EXL3 3.05 bpw, MTP on | DGX Spark recipes | Receipt ↗ |
| ~0.9s | Median voice turn to first audio, DGX-01, 2026-10-22 | LocalRouter | Receipt ↗ |
| 6.3s | Warm 1024×1024 Qwen-Image 2.1 Turbo generation, DGX-01 | LocalRouter | Receipt ↗ |