A field guide to local AI
Local intelligence.
Shared experience.
A DGX Spark, an ASUS GX10, an RTX 5090, and a Mac Studio. A working local AI lab, documented through configurations, measured experiments, and the lessons that changed our minds.
CONFIG VERIFIED OCT 23, 2026 / 45 PUBLISHED STUDY RECORDS
Coding + scout + native vision
DGX Spark + GX10
DeepSeek V4.1 Flash 1017 · EXL3 · normal
Selectors: coder · scout · vision
524,288 configured · 4 shared slots
Persistent TensorFold TP2 · 32,768 output cap · high reasoning default
Native and routed text/tools/image checks passed. Small matched coding panel tied GLM. Full-window retrieval and concurrent long-context serving remain unqualified. Dated configuration; moving dots illustrate requests.
GB10 systems
2 × 128 GB
DGX Spark + ASUS GX10. DeepSeek EXL3 through TensorFold TP2; four shared inference slots.
Media GPU
RTX 5090
32 GB GDDR7, Ryzen 9 9950X3D, Windows + WSL2.
Apple silicon
M5 Max · 48 GB
Mac Studio, 40 GPU cores. Agent, writing, embeddings, and local inference.
Fresh inspection / Oct 23, 2026
DeepSeek on two GB10 hosts. 4 shared inference slots.
DeepSeek V4.1 Flash 1017 now serves coder, scout and vision through TensorFold on the DGX Spark and ASUS GX10. The matched Hermes panel tied GLM on coding; short capacity probes, memory observations and cache replay have their own limits.
Open the DeepSeek setup guide ↗01 / Build your baseline
Find your starting point.
DGX Spark & GX10: capacity with discipline
Our DGX Spark and ASUS GX10 now run DeepSeek through TensorFold TP2. Understand unified memory, four shared slots, model staging and measured limits.
GUIDE 02 / 3 MIN READ
RTX 5090 on Windows + WSL2
Verify the Windows driver from Linux, enable CUDA, and budget a 32 GB GPU for useful work.
GUIDE 03 / 2 MIN READ
Mac Studio: MLX, Metal & memory ownership
Run local models on Apple silicon without confusing installed memory, configured context, and verified capacity.
GUIDE 04 / 2 MIN READ
02 / A working lab
Four machines. Deliberate roles.
NVIDIA DGX Spark
GB10 / 128 GB
Gateway coordinator + TP2 coder head
02 /ASUS Ascent GX10
GB10 / 128 GB
TP2 coder worker
03 /Mac Studio
M5 Max / 48 GB
Agent, writing, local images + persistent embeddings
04 /RTX 5090 workstation
RTX 5090 / 32 GB
Media, perception + on-demand specialists
Current placement is verified from active configuration. Older experiments keep their original hardware and dates.
03 / What the measurements taught us
Match the model to the job.
Historical / September Studio evaluation
Speed is only one axis.
Protocol & limitations ↗M5 Max · 40 GPU cores · 48 GB · Sep 22–23, 2026 · thinking off. Lightning here is the evaluated Nemotron artifact. The current SuperQwen alias has no score in this chart.
HumanEval / % ↑
Code decode / tok/s →
Coding fallback / MLX 4-bit
95.1%
HumanEval · 156 / 164 tasks passed
All 164 original tasks, one greedy chat sample, original executable tests, no repair or retries. Chat-adapted pass@1; training contamination is possible. Gemma was a writing-role evaluation and did not run this suite.
Speed: median of three short 512-output code probes on a loaded backend. Timing and quality use different prompts; this is not full-task wall time.
The experiment notebook
The fastest model
doesn’t win every job.
In September's Studio evaluation, Lightning decoded faster while Qwen passed more coding tests. Our Spark studies also include reversals, rejected contenders, and optimizations that only helped one workload.
See the experimentsMTP / Spark code
+57.4%
64.1 → 100.9 tok/s
Six draft positions, one quiet slot
MTP / Spark creative
−17.0%
64.6 → 53.6 tok/s
The same switch slowed writing
JUL 11, 2026 / HISTORICAL QWEN3.6 Q6_K / REPEAT SPREAD NOT RETAINED.
Jul → Oct 2026 / Learning in public
The setup changed. The lessons stayed.
Chapter 01 / Jul 11, 2026
+57.4% code / −17.0% creative
Six-position MTP changed two workloads in opposite directions. The optimization became a per-workload decision.
Read the decision ↗04 / The public journal
The decisions behind the setup.
DeepSeek on two GB10 hosts, with the coding gate still intact
Normal DeepSeek now serves four shared slots through TensorFold. The matched GLM panel tied on coding; capacity, cache replay, memory and maintenance have separate evidence.
Two GB10 nodes, eight slots, and the limits behind the settings
A fresh inspection of both workers, October runtime probes, cache capacity, startup cost, and the Studio alias that changed models.
Publishing the benchmark methods and model history
The lab’s test prompts, scoring rules, promotion criteria and model variants, including a complete Aider run and an incomplete Flash archive.
Agentic repo prompt / Illustrated process
Give the agent a loop.
Keep the evidence.
A bounded improvement becomes a verified change, an independent review decision, and a durable record. The model's decode score measures one small piece of that process.
01 / Inspect
Read the repository instructions and evidence. Choose one useful problem inside the authorized scope.
QABot returns MERGE or HOLD. MergeBot handles eligible squash merges. Authorization and a useful next pass bound the loop.