Flash-Next passed all 24 attempts on two Arc Pro B60s with XeStrata

The model took 13 minutes to load from our hard-drive archive. Once warm, the median passing reply took 2.70 seconds.

If you want an Arc PC to fix a Python function, pull fields out of a document or choose a tool call, getting a 125-billion-parameter model to load is only half the job. You still need to know whether its answers hold up and how long you’ll wait for them.

XeStrata running Qwen3.8-Flash-Next passed all 24 attempts across 12 short tasks on our two Intel Arc Pro B60s. The median passing reply took 2.70 seconds, including the wait for output to start. The scored run took 13 minutes 16 seconds to load from our storage setup. To see if the setup fits your work, time startup separately and check warm replies against the tasks you plan to run.

What Flash-Next got right

We sent four coding fixes, four document extractions and four tool decisions, then repeated the set in the same order. Each request had a 512-output-token cap. A coding reply had to preserve the function signature and pass its tests. An extraction had to return exactly the requested fields and values. For tool decisions, the action and arguments both had to match. In one case, the correct decision was to do nothing. The tasks were synthetic and the tool decisions stayed in the test.

Two attempts per case, one request at a time
TaskPassing attemptsMedian reply timeObserved range
Coding fixes8 of 83.86 s1.91–11.93 s
Document extraction8 of 82.78 s2.61–3.00 s
Tool decisions8 of 82.11 s1.13–2.73 s

Reply time runs from the request to the end of the streamed answer. The coding tests exercised four small functions and ran in an isolated CPU sandbox after the GPU run. The results cover these short cases. Larger patches, longer documents and tools that can act on real systems need their own tests.

What it took to get both cards ready

The first startup hit its 12-minute limit while the second GPU’s expert cache was still reading from the hard-drive mirror. With a longer deadline, the same profile loaded in 12 minutes 54 seconds for a smoke test and 13 minutes 16 seconds for the scored run. The model shards stayed on the mirrored ZFS pool. The embedding table used direct reads from an NVMe copy. That storage layout matters when comparing load times.

Peak sampled resident GPU memory was 19.58 and 18.86 GiB on the two cards. The engine’s peak resident system memory was 2.53 GiB. The container’s peak memory charge, which includes charged file cache, was 4.49 GiB. Sampled container swap use stayed at zero.

We used XeStrata xe0.2.2-1, the original ISTA-DASLab IQ2_XS shards, an 8,192-token context limit and a split at layer 25. Experts were memory-mapped from the model files during loading. The engine used its own Intel XMX kernels and ran with the model’s MTP draft layer disabled. A single-card setup needs a separate fit and performance check. XeStrata release · Pinned model files

How to check your own setup

Download the Linux release and verify the package with sha256sum -c SHA256SUMS --ignore-missing in the download directory. The release DEBs require glibc 2.43 or newer. Check yours with getconf GNU_LIBC_VERSION. Our Ubuntu 24.04 host had 2.39, so we used the release RPM in a Fedora container. Include the Intel Level Zero driver and loader in the container, then run xestrata --check inside it to list the devices the engine can use. Pinned container instructions

Keep the original GGUF shards together and point setup at their directory with --gguf-dir. Use --family qwen --model IQ2_XS --context 8192 --parallel 1 and choose the card numbers from --check. For the two-card split, use --gpus with those two numbers and --layer-split 25. If setup enables MTP, remove the --mtp option and its path from the test configuration’s engine arguments and set --spec 2 to match this run. Keep the server on loopback. GPU selection and split settings

Start timing when you launch the server. Set PORT to your test port. Stop the startup clock when the health response shows loaded is true and confirms the model and context. The second command should return 42 in the assistant message. Then repeat a few tasks from your own workload and check the answer each time. Count replies you have to redo as failures.

PORT=8188
curl --fail --silent --show-error --max-time 5 \
  "http://127.0.0.1:$PORT/health" | python3 -m json.tool
curl --fail --silent --show-error --max-time 90 \
  -H 'Content-Type: application/json' \
  --data '{"model":"qwen3.8-flash-next-iq2_xs","messages":[{"role":"user","content":"What is 17 plus 25? Reply with the decimal integer only."}],"temperature":0,"max_tokens":32,"reasoning_effort":"none","chat_template_kwargs":{"enable_thinking":false}}' \
  "http://127.0.0.1:$PORT/v1/chat/completions" | python3 -m json.tool

Download the per-attempt results and resource summary for the exact times, verdicts, fixture hashes and memory counters.