Skip to content
← All setup guides

Guide 09 / Deployment

DeepSeek on two GB10 hosts: stage, validate, compare

The October 23 DeepSeek deployment, from verified checkpoint and Engram staging to four shared slots, a matched GLM control and reversible operations.

8 MIN READ / REVIEWED Oct 23, 2026

01What is running

Normal DeepSeek V4.1 Flash 1017, EXL3 EAQ 2.9 bpw, runs on our DGX Spark and ASUS GX10 through the pinned TensorFold native Zig engine. Each host has 128 GB unified memory. Coder, scout and vision share four slots. The optional abliteration overlay is archived, not applied.

Inspected local deployment policy · not a launch script

# Inspected deployment policy · 23 Oct 2026
# Inventory excerpt; not a standalone launch command
model: DeepSeek-V4.1-Flash-1017
quantization: EXL3 EAQ 2.9 bpw
runtime: TensorFold native Zig
nodes: 2
tensor_parallel_size: 2
mode: normal
ABLIT: 0
CONTEXT: 524288
PARALLEL: 4
MAX_TOKENS: 32768
THINKING: 1
REASONING_EFFORT: high
MIN_AVAIL_GIB: 112
MEMGUARD_GIB: 2.5
MEMGUARD_HARD_GIB: 1
MEMGUARD_HOLD_S: 4
shared_roles: [coder, scout, vision]

02Stage the complete model before making space

The verified archive includes the EXL3 checkpoint and separate Engram inputs, totaling 413,743,842,936 bytes. Download to the model archive first, verify each manifest-listed file, then stage and recheck the required local inputs. Archive-verified superseded models can make room without losing the rollback.

Keep Engram inputs on local storage for both ranks. Preserve the incumbent’s image, checkpoint, configuration and spill resources. A free-disk figure does not establish unified-memory headroom.

03Make preparation repeatable

Pin the recipe, native engine, input tools and image. Our installed recipe revision is 370daf0, engine 629c524 and patch set cc4781a03264. Both ranks use the same image. The worker’s second RoCE rail was repaired during installation, and connectivity was checked before runtime startup.

Preparation and runtime are separate supervised steps. A missing CLI path was corrected without changing the model or build pins. Preserve failure receipts and resume from verified inputs.

04Validate native serving and every gateway hop

These checks passed during installation, including a 21,875-position prompt. The 524,288 setting and four slots remain configuration; full-window retrieval and concurrent long-context capacity were not qualified.

  • Check the exact response model, not just the coder alias.
  • Test text, streaming completion and usage, structured tools and tool-result roundtrip.
  • Exercise image input, thinking on/off, output limits and stop behavior.
  • Verify head, worker and HTTPS entry points reach the same native model.
  • Record sampled head and worker memory separately.

05Keep model comparison separate from tuning

The frozen primary Hermes comparison used temperature 0.2, thinking off and 4,096 output positions per request for both models. Both accepted 1/1 routine coding tasks and 0/3 hard coding tasks. DeepSeek correctly rejected a defective change that GLM approved; both handled the admission and healthy-review controls correctly.

Production then selected high reasoning and the native 32,768 allowance, four shared leases and a four-worker target. Short non-thinking inspection and planning have separate budgets. Changed settings do not turn the original coding tie into a demonstrated model advantage.

06Resume held work with its history intact

A copied 63,164-position history produced valid structured responses in 43.9 s cold and 8.8 s warm, with 61,440 cached positions on the second request. No tools were executed in that replay. The real held workers kept their original sessions and frozen prompts.

After regression checks and a supervised restart, the loop resumed at 03:57 CDT on October 23. Fresh work uses the DeepSeek prompt revision. Count verified draft changes, independent review, blockers and merges as distinct outcomes; held runs can include GLM work.

07Keep maintenance status honest

sparkDash 2.70.0 is installed; customized llama-swap v257, NVIDIA driver 580.173.02 and the inspected OS remain in place. Available gateway, platform and monitoring updates are a separate maintenance plan. The software stack page records installed and available versions with their observation dates.

Primary references

Upstream docs describe supported behavior. Lab-specific configurations and numeric results retain their inspection or experiment dates.

Keep going

Related guides