Skip to content

Hardware / verified Oct 23, 2026

A lab built around the workload.

Two GB10 hosts for the primary coder. Apple silicon for agents, writing, and embeddings. A discrete GPU for media and specialists. Every machine has a job and a memory budget.

Inside the local labIllustrated flow / Oct 23, 2026
Coding + scout + native vision routing through llama-swapDeepSeek V4.1 Flash 1017 · EXL3 · normal routes to DGX Spark + GX10. The two GB10 hosts share an explicitly configured tensor-parallel coder. Every other route uses an individual host. Moving dots illustrate requests, not live activity.DGX SparkGB10 · 128 GBASUS GX10GB10 · 128 GBMac StudioM5 Max · 48 GBRTX 509032 GB GDDR7llama-swapROUTING + RESIDENCYTWO NODES · TENSORFOLD TP2 · FOUR SHARED SLOTS

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.

Evidence before labels

DeepSeek on two GB10 hosts: four shared slots, with the evidence attached

October 23 review covers the running GB10 deployment, coordinator selectors, Studio launchers and passive inventory, plus workstation hardware, launcher flags and passive inventory. llama-swap routes requests; TensorFold separately distributes DeepSeek across two GB10 hosts. Memory and lifecycle policies remain local to each placement.

  • Both running GB10 containers, installed settings and gateway selectors · October 23
  • Studio hardware, managed launchers and passive gateway inventory; workstation hardware, launch flags and gateway inventory · October 23
  • Exact native identity, text, streaming, tools and image validation · October 23
  • Frozen matched Hermes controls · October 22–23; one routine and three hard coding tasks per model
  • One-hour memory observation, 1/2/4-request capacity probes and a held-history replay · October 23
  • sparkDash update and separate pending-maintenance inventory · October 22
Read the configuration audit ↗

GB10 pair / Reviewed Oct 23, 2026

DeepSeek is serving. The coding panel is a tie.

Normal DeepSeek V4.1 Flash 1017 runs through TensorFold TP2 on the DGX Spark and ASUS GX10. The matched GLM comparison, short serving probes and memory observation answer different questions.

Shared inference slots
4

Coder, scout and vision share this capacity. Studio agent and embeddings keep independent routes.

Configured context
524,288

A serving limit. Local full-window retrieval and concurrent long-context capacity remain unqualified.

Initial memory minima / GiB
4.93 / 6.55

Head / worker available memory across one hour, 1,800 samples each. A bounded observation; guards remain active.

October 22–23 · Same frozen primary Hermes protocol, temperature 0.2, thinking off, 4,096 output allowance. Coding handoffs and diagnostic decisions keep separate denominators.
Model / dateRoutine codingHard codingDefective reviewOther controls
GLM · abliterated1 / 10 / 3Incorrect approvalAdmission 1 / 1
Healthy review 1 / 1
DeepSeek · normal1 / 10 / 3Correct rejectionAdmission 1 / 1
Healthy review 1 / 1

One routine task, three hard tasks, one visit each, one harness. Behavior, scope, report, exact identity and cleanup are required. This local panel is not an official SWE-bench score. Later production high-reasoning tuning is separate; no difficult-coding advantage was demonstrated.

Memory / Four separate budgets

Capacity is local.
Placement is deliberate.

The gateway selects a backend. Only the two GB10 hosts run an explicitly distributed coder. Routing the Studio and 5090 does not turn their memory into one model allocation.

Weights + KV cache + activations + runtime + headroom all share the usable budget. Quantization changes weights; context and concurrency change cache pressure.

Mac Studio

48 GB installed unified memory, 18 CPU cores. The September evaluation recorded a Metal recommended working set of 37.44 GiB; 48 GB is not an unrestricted model budget.

Budget & setup guide ↗

01 / GB10 / 128 GBVerified

NVIDIA DGX Spark

Gateway coordinator + TP2 coder head

Compute
NVIDIA GB10 · Blackwell
Memory
128 GB unified memory. GPU memory.total reports N/A on this platform; budget from system memory and actual runtime allocations.
Why this role
Coordinator gateway and TensorFold head for normal DeepSeek V4.1 Flash 1017 EXL3. Four shared coder/scout/vision slots, 524,288 configured context, high reasoning default. October 23 memory and capacity probes are bounded observations; GLM rollback is preserved.
Setup guide ↗

02 / GB10 / 128 GBVerified

ASUS Ascent GX10

TP2 coder worker

Compute
NVIDIA GB10 · Blackwell
Memory
128 GB unified memory class; Linux reports about 121.6 GiB of system memory. This is a distinct host, not a second DGX-branded Spark.
Why this role
ASUS GX10 supplies rank 1 of the DeepSeek TensorFold TP2 deployment. Both GB10 nodes run the same pinned image and still report driver 580.173.02. The worker’s second RoCE rail was repaired during installation; no bandwidth result is claimed.
Setup guide ↗

03 / M5 Max / 48 GBVerified

Mac Studio

Agent, writing, local images + persistent embeddings

Compute
Apple M5 Max · 40 GPU cores
Memory
48 GB installed unified memory, 18 CPU cores. The September evaluation recorded a Metal recommended working set of 37.44 GiB; 48 GB is not an unrestricted model budget.
Why this role
M5 Max and 48 GB memory rechecked October 23. SuperQwen MLX 4-bit serves lightning/agent; Qwen coding fallback, Gemma writer/native vision and MFLUX image alternatives share one exclusive primary group. Qwen embeddings have an independent persistent group. Only the embedder was ready during the passive check; configured roles load on demand. September Nemotron Lightning results remain historical.
Setup guide ↗

04 / RTX 5090 / 32 GBVerified

RTX 5090 workstation

Media, perception + on-demand specialists

Compute
NVIDIA GeForce RTX 5090 · 32 GB GDDR7
Memory
Driver reports 32,607 MiB GPU memory. WSL guest reports about 94.3 GiB system RAM; that is the guest allocation, not a verified Windows installed-memory figure.
Why this role
AMD Ryzen 9 9950X3D, Windows + WSL2, NVIDIA driver 617.42. Image, video, music, OCR and translation use model-specific GPU pipelines; Parakeet transcription and Kokoro speech use CPU. No model backends were loaded at the passive October 23 check. The peer uses model TTLs and legacy matrix sets covering some older combinations. Heavy-pipeline concurrency was not exercised. Coding and Omni alternatives are inventory, with no new speed or quality result claimed.
Setup guide ↗

Current placement

Follow the role to the backend.

Software stack
Configured model roles, placement and limits as verified Oct 23, 2026; individual hardware observations retain their dates
Role / modelPlacementLimits / lifecycle

Coding + scout + native vision

DeepSeek V4.1 Flash 1017 · EXL3 · normal

DGX Spark + GX10

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.

Studio agent

SuperQwen3.8-27b-abliterated · MLX 4-bit

Mac Studio

1 active decode/prompt request · 2 GB prompt-cache budget

Exclusive primary swap group · TTL 600 s

Managed SuperQwen launcher and Studio inventory inspected October 23: thinking off, one decode/prompt request and a 2 GB prompt cache. Studio-local scout is also this process; the coordinator scout selector goes to DeepSeek. September Nemotron Lightning scores describe a different artifact.

Local coding fallback

Qwen3.8-27B · MLX 4-bit

Mac Studio

64K budget · 1 active request

Exclusive primary swap group · TTL 900 s

The coordinator retains the legacy coder-spark alias for this Studio fallback. September Qwen HumanEval and custom-suite evidence remains separately dated.

Writing / vision agent

Gemma-4 26B-A4B QAT · Q4_K_M

Mac Studio

64K · separate writer and native-vision configs

Exclusive primary swap group · TTL 900 s

Writer is text-only; agent-gemma uses a matching BF16 vision projector. Each is one request and they swap exclusively. The published September writing evaluation used the writer configuration.

Embeddings

Qwen3-Embedding-4B · Q8_0

Mac Studio

8,192 configured

Independent persistent group

Persistent is a lifecycle policy; loaded state is a separate observation.

Images / video / music

Z-Image Turbo / base / portrait / uncensored, Krea 2, Qwen Image 2.1 / Turbo / edit, FLUX.2 klein 4B, Wan 2.2, LTX 2.3, MiniMax H3, Hunyuan 1.5, Lingbot, ACE-Step, MiniMax Music 3

RTX 5090

Pipeline-specific resource budget

On demand · model TTLs 300–1,800 s · shared 32 GB GPU budget

Coordinator selectors, launch flags and the legacy workstation matrix were inspected October 23. Named matrix sets cover some older combinations; several newer pipelines are outside those sets. Heavy-pipeline concurrency was not exercised, and route presence adds no speed or quality score.

OCR / translation / speech

Qwen3.5 9B Q6_K OCR / translation; Parakeet TDT 0.6B v3 ONNX INT8 transcription; Kokoro-82M speech

RTX 5090

OCR / translation: 65,536 configured · 1 request per route

On demand · TTL 600 s · OCR / translation use GPU; Parakeet / Kokoro use CPU

Workstation configuration inspected October 23. CPU speech routes are distinct from the GPU media budget. The coordinator vision selector now goes to DeepSeek on the GB10 pair.

Studio-local image alternatives

Z-Image Turbo / Krea 2 Turbo · MFLUX 8-bit

Mac Studio

1 request · 1,024 px · 8 steps by default

Exclusive primary swap group · TTL 1,800 s

These namespaced Studio alternatives share the large-model group. The coordinator image, z-image and krea-2 selectors target the RTX 5090 instead. Configuration inspected October 23; no new speed or quality result is claimed.

Alternate coding / perception inventory

Nemotron 3 Nano Omni 30B UD-Q4_K_M; Qwen3-Coder-Next 80B UD-Q4_K_XL with CPU MoE offload; Ornith 1.0 35B Q5_K_M

RTX 5090

Model-specific context and sequence budgets

Configured alternatives · TTL 0 for Omni; 300 s for the two coding candidates

Legacy coder-dgx now targets the workstation, not the DGX. A zero TTL does not establish residency or persistence. These selectors are inventory, not evidence that a candidate outperforms the primary DeepSeek coder.