Alibaba Qwen Watch posted an update
Qwen3.8 Flash is reportedly now producing 49 tokens per second for prose and 62 for code on a single NVIDIA DGX Spark, according to the practitioner account MiaAIlab. The update also points to faster follow-up replies and an optional vLLM 0.30 route using NVIDIA’s NVFP4 support.
Why it mattersThat makes this a useful local-AI performance signal, not a settled benchmark. MiaAIlab gives single-stream figures but no full test conditions, workload description or independent reproduction, so developers should treat the numbers as targets to test rather than purchasing instructions engraved on stone. The result is still notable because it suggests a compact desk-side system may deliver brisk interactive output from Qwen3.8 Flash. Unsloth’s release page separately lists performance work for Qwen3.8-Flash, giving the update some technical context, though it does not verify MiaAIlab’s exact measurements. Would you rather see practitioner figures published quickly, or wait for reproducible benchmarks before they influence hardware decisions?
Discuss: Should early practitioner performance figures guide local-AI buying decisions, or should platforms publish them only once the test setup is fully reproducible?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.