# Strix Halo or DGX Spark for local AI? What we measured on both boxes

> The most searched comparison in the local-AI mini-box class is answered everywhere by spec sheets and nowhere by one lab measuring both. We own both platforms. Here is what each one demonstrably did, which workload belongs on which box, and the comparison we refuse to fake.

- Published: 2026-08-20
- Evidence level: lab_repeated
- Funding: Self-funded internal research
- Canonical: https://agmind.ai/reports/strix-halo-vs-dgx-spark/

The honest headline first: these are different tools that the market
insists on cross-shopping. One box is a measured, boring workhorse for the
model class most people actually run. The pair of the other is the only
desk-format way we know to serve the largest open MoE class — with real
operational sharp edges. We own and measured both; nobody in this
comparison's search results does.

## Why is this comparison so hard to find measured?

Because it is genuinely awkward to do honestly. The two platforms want
different runtimes, favor different model classes, and every published
head-to-head we found either reruns vendor numbers or quotes two
incompatible community figures side by side —
[which is how benchmark folklore is born](https://agmind.ai/essays/benchmark-number-config-in-disguise/).
What follows keeps the two boxes in separate, clearly-scoped columns, and
the last section names the measurement that would join them properly.

## What the Strix Halo box actually did

One Ryzen AI Max+ 395 unit with 128 GB unified memory, llama.cpp pinned by
digest, serving a 35B-class MoE. Everyday requests: first answer token in
**210 ms** ([strix.qwen36.interactive2.c1.ttfa-nothink](https://agmind.ai/claims/strix.qwen36.interactive2.c1.ttfa-nothink/)). Four
simultaneous chats:
**338 ms** ([strix.qwen36.interactive2.c4.ttfa-nothink](https://agmind.ai/claims/strix.qwen36.interactive2.c4.ttfa-nothink/)) to the
first word, completion holding at
**100.0 % of requests** ([strix.qwen36.interactive2.c8.completion-nothink](https://agmind.ai/claims/strix.qwen36.interactive2.c8.completion-nothink/)) even
at eight streams. A 32k-token document costs
**33965 ms** ([strix.qwen36.longctx.c1.ttft-32k-en](https://agmind.ai/claims/strix.qwen36.longctx.c1.ttft-32k-en/)) up front, and
three continuous hours of serving held pace at
**30.0 ms/token** ([strix.qwen36.endurance.c4.itl-median](https://agmind.ai/claims/strix.qwen36.endurance.c4.itl-median/)). The full tables
are in [the benchmark repository](https://github.com/botAGI/strix-halo-llm-benchmarks)
and [the 128 GB report](https://agmind.ai/reports/what-128gb-unified-memory-runs/).

## What the DGX Spark pair actually did

Two GB10 units over a 200G link served DeepSeek-V4-Flash — a
284B-parameter MoE — through vLLM, stably enough to publish
[the raw runs](https://github.com/botAGI/dspark-0731-gb10) and
[the recipe with its traps marked](https://agmind.ai/reports/deepseek-v4-flash-0731-dgx-spark-recipe/).
The pair walked a context curve to a million tokens and
[told the truth about the bill](https://agmind.ai/reports/dspark-speculative-1m-dgx-spark/):
decode plateaus at depth, but the first fill of a million-token window is
an event you schedule, not a request you fire. Bring-up was real work — a
[kernel-selection flag](https://agmind.ai/reports/dsv4-flash-moe-backend-dgx-spark/) the
automatic mode skips separates the fast configuration from the one most
people measure, and the stock GPU monitoring
[half-works at best](https://agmind.ai/reports/dgx-spark-gpu-monitoring/). The buyer's
answer, including who should not buy one, is
[its own page](https://agmind.ai/reports/dgx-spark-worth-it/).

## Which box for which workload?

From what we measured, not from spec sheets. A personal assistant, a small
team's chat box, document Q&A in the tens-of-thousands-of-tokens range, JSON
automation: the Strix Halo box does all of it today, quietly, with the
numbers above. Serving the largest open MoE class under your own roof, or
long-context research where a million-token window amortizes: that is what
the Spark pair demonstrably did, and nothing in our fleet at this size
matched it. If your question is "which one for a home assistant", you are
choosing between a measured yes and an overqualified maybe. If your
question is "can I serve the frontier open models locally", only one of
these platforms did it in our lab — as a pair, with seams.

## Why don't our numbers match the reviews you read?

Because most published figures for both platforms are configs in disguise:
unstated workload profiles, moved defaults, pre-update firmware. We wrote
[a reconciliation of the disagreeing Spark numbers](https://agmind.ai/reports/dspark-speed-numbers-disagree/)
instead of adding another orphan figure, and
[a protocol](https://agmind.ai/reports/how-to-benchmark-local-llm/) for publishing numbers
that can be disagreed with. Both apply to any figure on this page.

## What we have NOT measured

The clean comparison — same model, same runtime, same workload, both
boxes, one harness — does not exist yet, here or anywhere we could find.
Our Strix lane ran Qwen3.6 and gemma-4 under llama.cpp; our Spark lane ran
DeepSeek-V4-Flash under vLLM. Any table putting those numbers in adjacent
columns would be manufacturing the exact folklore we exist to debunk, so
this page does not contain one. The same-harness cross-device cell is
planned; when it lands, its claims will carry system-level comparison
scope and this page will be superseded by something better. Until then:
two honest columns beat one fake table.

Every Strix figure above has a permanent page with raw runs in
[the claim registry](https://agmind.ai/claims/); every Spark statement links its published
run repository.

---

Machine-readable claim registry: https://agmind.ai/claims.json · llms.txt: https://agmind.ai/llms.txt