koalfied-coder claims a reproducible two-DGX-Spark setup runs DeepSeek Flash v4 at a sustained 67–84 tokens per second with fast prompt evaluation, making the model practically usable for high-throughput local inference.

state: expiredheat: lowuncertainty: highknownscott: mediumdeepseek local-inference inference-economicskoalfied-codertonyd2wildNVIDIAASUS

What is this?

A community-built deployment runs DeepSeek-V4-Flash across two NVIDIA DGX Spark/GB10 systems using a development build of vLLM, with reports of roughly 60–67 tok/s sustained single-stream decoding and peaks near 84 tok/s; concurrency claims reach roughly 220 aggregate tok/s. The supplied snippets include configuration details and independent engineering discussion, but they do not firmly establish koalfied-coder’s identity or ASUS’s role, and results conflict: another dual-Spark test reports stable sustained generation of only about 41 tok/s. The headline performance is therefore plausible and partly reproducible, but not uniformly confirmed across workloads and configurations.

Why it matters to Scott

The radar already tracks this exact development in `radar:deepseek-v4-twinspark-dgx-serving`. Its disputed dual-node throughput bears directly on Scott’s hardware-aware local-inference work and could inform whether scaling his self-hosted model substrate is operationally and economically worthwhile, but the conflicting results do not yet establish a reliable new conclusion.

dev:concept.hardware-aware-local-inferencedev:project.gamepcip:concept.ai-unit-economicsradar:deepseek-v4-twinspark-dgx-servingradar:concept.local-inferenceradar:concept.distributed-inferenceradar:concept.inference-economics

queries asked of Scott's wikis

local inference throughput economics versus API inference
dual-node unified-memory model serving
local coding-agent inference latency requirements
open-model quantization and speculative decoding
high-throughput local inference concurrency
DGX Spark or GB10 deployment experience

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (2) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit ⭐	67-84 t/s DeepSeek flash v4 off 2x GX10s LocalLLaMA	koalfied-coder	45	28
🟠 reddit	Humaneval benchmark for Deepseek V4 Flash 0731 vs GLM5.3 Flash on 2x DGX Spark setup LocalLLaMA	serige	23	11

Interpretation history

2026-09-02T16:46:10Z
After repeated checks, no matched reproduction, logs, or coding-agent benchmark has narrowed the throughput discrepancy, and no confirming evidence is now expected within the episode’s horizon. Preserve dual-Spark operability as credible, but let the specific 67–84 tok/s and inference-economics claim fade pending a genuinely new reproduction.
2026-08-31T16:37:57Z
The refreshed discussion adds no matched reproduction, logs, or coding-agent benchmark, so it does not narrow the gap between the headline 67–84 tok/s claim and the slower independent result. Dual-Spark operability remains credible, but dependable throughput and inference economics remain unsettled.
2026-08-30T16:31:35Z
The refreshed comments add no matched configuration, logs, or workload-relevant benchmark, so they do not resolve the gap between the reported 67–84 tok/s and the slower independent result. Dual-Spark operability remains credible, while headline throughput and economics stay unsettled.
2026-08-30T15:34:36Z
The refreshed discussion still supplies no matched configuration, logs, coding benchmark, or independent reproduction of the claimed 67–84 tok/s range. Dual-Spark operability remains credible, but the headline throughput and inference-economics conclusion are unchanged and unsettled.
2026-08-30T05:31:35Z
The refreshed discussion adds only an unsupported claim that GLM 5.3 Flash is preferable, not a matched reproduction or new benchmark evidence. DeepSeek V4 Flash is credibly operable on two DGX Sparks, but the claimed 67–84 tok/s performance and resulting inference economics remain unresolved.
2026-08-30T01:24:03Z
Refreshed comments add requests for coding benchmarks and alternative quantization tests, but no matched configuration, logs, or reproduction of the claimed 67–84 tok/s range. The setup’s operability remains credible while its headline throughput and practical economics remain unsettled.
2026-08-29T23:25:54Z
The case has moved beyond a single-operator anecdote because a second user independently runs and benchmarks the same model on two DGX Sparks. However, the available results do not reproduce the headline 67–84 tok/s and instead reinforce that throughput is configuration- and workload-sensitive, so the practical-performance claim remains unsettled.
2026-08-29T23:22:37Z
evidence attached: reddit.post.1w215qm — Independent two-DGX-Spark benchmarking directly tests the case’s claim about practical local throughput and compares it with a competing model.
2026-08-29T19:42:03Z
No substantive evidence arrived: modest engagement growth without new comments, logs, or independent reproduction leaves the conflicting throughput claim unresolved. The setup remains worth tracking, but this delta adds no confidence or urgency.
2026-08-29T19:32:33Z
grounded: known/medium — The radar already tracks this exact development in `radar:deepseek-v4-twinspark-dgx-serving`. Its disputed dual-node throughput bears directly on Scott’s hardwa
2026-08-29T19:27:10Z
case created — This is a concrete performance claim tied to a public setup repository, but it currently has only one operator report.

Decision trace

09-03 02:46expireAfter repeated checks, no matched reproduction, logs, or coding-agent benchmark has narrowed the throughput discrepancy, and no confirming evidence is now expected within the episode’s horizon. Preser
09-03 02:46alert_silentThe staleness trigger carries no new consequential evidence; any future matched benchmark can reopen the claim as a fresh delta.
09-03 02:46alert_routeThe staleness trigger carries no new consequential evidence; any future matched benchmark can reopen the claim as a fresh delta.
09-01 02:37repriceThe refreshed discussion adds no matched reproduction, logs, or coding-agent benchmark, so it does not narrow the gap between the headline 67–84 tok/s claim and the slower independent result. Dual-Spa
09-01 02:37alert_silentThis delta is engagement and repetitive discussion rather than consequential new evidence; it can wait for a matched configuration reproduction, detailed performance logs, or workload-relevant benchma
09-01 02:37alert_routeThis delta is engagement and repetitive discussion rather than consequential new evidence; it can wait for a matched configuration reproduction, detailed performance logs, or workload-relevant benchma
08-31 02:31repriceThe refreshed comments add no matched configuration, logs, or workload-relevant benchmark, so they do not resolve the gap between the reported 67–84 tok/s and the slower independent result. Dual-Spark
08-31 02:31alert_silentThis is repetitive discussion rather than consequential new evidence and can wait for a matched reproduction, detailed logs, or coding-agent benchmark.
08-31 02:31alert_routeThis is repetitive discussion rather than consequential new evidence and can wait for a matched reproduction, detailed logs, or coding-agent benchmark.
08-31 02:21sensor_dirtycomment_update
08-31 01:34repriceThe refreshed discussion still supplies no matched configuration, logs, coding benchmark, or independent reproduction of the claimed 67–84 tok/s range. Dual-Spark operability remains credible, but the
08-31 01:34alert_silentThis delta is repetitive discussion rather than consequential evidence, so it can wait for a matched reproduction or workload-relevant benchmark.
08-31 01:34alert_routeThis delta is repetitive discussion rather than consequential evidence, so it can wait for a matched reproduction or workload-relevant benchmark.
08-31 01:21sensor_dirtycomment_update
08-30 21:21sensor_dirtyengagement_update
08-30 19:21sensor_dirtyengagement_update
08-30 19:21sensor_dirtyengagement_update
08-30 15:31repriceThe refreshed discussion adds only an unsupported claim that GLM 5.3 Flash is preferable, not a matched reproduction or new benchmark evidence. DeepSeek V4 Flash is credibly operable on two DGX Sparks
08-30 15:31alert_silentNo consequential new evidence arrived; the comment refresh neither validates the disputed throughput nor changes a hardware decision for Scott. It can wait for matched configurations, logs, or coding
08-30 15:31alert_routeNo consequential new evidence arrived; the comment refresh neither validates the disputed throughput nor changes a hardware decision for Scott. It can wait for matched configurations, logs, or coding
08-30 15:21sensor_dirtycomment_update
08-30 11:24repriceRefreshed comments add requests for coding benchmarks and alternative quantization tests, but no matched configuration, logs, or reproduction of the claimed 67–84 tok/s range. The setup’s operability
08-30 11:24alert_silentThe new discussion is exploratory rather than evidentiary and does not change the hardware decision signal; it can wait for a matched benchmark or independent reproduction of the claimed throughput.
08-30 11:24alert_routeThe new discussion is exploratory rather than evidentiary and does not change the hardware decision signal; it can wait for a matched benchmark or independent reproduction of the claimed throughput.
08-30 11:21sensor_dirtycomment_update
08-30 11:21sensor_dirtycomment_update
08-30 10:21sensor_dirtyengagement_update
08-30 09:25repriceThe case has moved beyond a single-operator anecdote because a second user independently runs and benchmarks the same model on two DGX Sparks. However, the available results do not reproduce the headl
08-30 09:25alert_silentIndependent operation makes the setup more credible, but the new evidence neither validates the disputed throughput range nor yields a clear hardware or inference-economics decision for Scott. It can
08-30 09:25alert_routeIndependent operation makes the setup more credible, but the new evidence neither validates the disputed throughput range nor yields a clear hardware or inference-economics decision for Scott. It can
08-30 09:22alert_silentA second user reports running DeepSeek V4 Flash on two DGX Sparks and names a reproducible recipe, but the supplied benchmark table is truncated before any quality, runtime, or stream-speed results. I
08-30 09:22surface_candidateA second user reports running DeepSeek V4 Flash on two DGX Sparks and names a reproducible recipe, but the supplied benchmark table is truncated before any quality, runtime, or stream-speed results. I
08-30 09:22alert_routeA second user reports running DeepSeek V4 Flash on two DGX Sparks and names a reproducible recipe, but the supplied benchmark table is truncated before any quality, runtime, or stream-speed results. I
08-30 09:22attachIndependent two-DGX-Spark benchmarking directly tests the case’s claim about practical local throughput and compares it with a competing model.
08-30 09:22propose_attachIndependent two-DGX-Spark benchmarking directly tests the case’s claim about practical local throughput and compares it with a competing model.
08-30 08:21sensor_dirtycomment_update
08-30 06:21sensor_dirtycomment_update
08-30 05:42repriceNo substantive evidence arrived: modest engagement growth without new comments, logs, or independent reproduction leaves the conflicting throughput claim unresolved. The setup remains worth tracking,
08-30 05:42alert_silentThe only change is engagement on the existing single-operator report; it neither validates the benchmark nor changes Scott’s local-inference decisions. Wait for an independent reproduction, benchmark
08-30 05:42alert_routeThe only change is engagement on the existing single-operator report; it neither validates the benchmark nor changes Scott’s local-inference decisions. Wait for an independent reproduction, benchmark