[Screenshot_20260719_140433_Firefox] 2+ trillion, about to open weight. Maybe a small companion but not very likely imo. [Screenshot_20260719_140719_Firefox]

swinging around many “facts”, i mean even talking about a bubble is wide accepted speculation but ok

Just running the same benchmark made by @technigma.ai , if ends fast ill share here too. For now during utilization.. 74.89 GB

I got similar speeds with Qwen 3.8 NVFP4. It runs about 15 output tokens/sec for one request and 14 tokens/sec per request at four-way concurrency

Has anyone tried NVFP4 speed over FP8?

Should be relatively easy to draft dflash/dspark head from it given that support is already there in inference. I get 17 t/s with no mtp on m5 max, 3.6 27b with mtp (not dspark) runs at 27-28 t/s, fp8.

Not a great benchmark at all, but with a few edits to change the target model of the Eugr Qwen3.5 35B A3B NVFP4 recipe to the Q 3.8 FP8, the performance on 1x spark starts at ~15.3tk/s, then after an intense dead-code cleanup of an existing cross domain component in a full stack email templating app…

unsloth Qwen3.8-27B-NVFP4 [image] unsloth/Qwen3.8-27B-NVFP4 · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

SGLang reportedly achieved 38 tok/s on DGX Spark with the following quant + DSpark head. [image] RadixArk/Qwen3.8-27B-NVFP4 · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science. […

╭─────────────────────────── 🏆 Benchmark Complete ────────────────────────────╮ │ │ │ Model: qwen-27 │ │ Score: 88 / 100 …

Qwen 3.8 is about to launch, open weight too

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10

MhaWay August 14, 2026, 5:24pm 82

Im curious about dflash/dspark generation speed actually

Topic		Replies	Views
Qwen3.8 27b - Upcoming release countdown DGX Spark / GB10 Projects	59	7667	August 17, 2026
Qwen3.8-27B-MixedInt4-AutoRound — Optimized for a Single DGX Spark DGX Spark / GB10 cuda , llama	42	7002	August 27, 2026
Single DGX-Spark - Qwen 3.8-Flash-Next at ~43tok/sec in Coding DGX Spark / GB10	75	14865	September 30, 2026
Qwen3.8-Flash-Next DGX Spark / GB10	316	36424	September 25, 2026
Qwen3.6-27B is out! DGX Spark / GB10 agentic-ai	300	34914	June 30, 2026
Qwen3.8-Flash-Next 180B, Single Solo DGX Spark With HashK-PLE NVFP4 DGX Spark / GB10 Projects	42	6703	September 5, 2026
Qwen3.8-27B-NVFP4 on a single DGX Spark — up to 1M context, vLLM+MTP measurements DGX Spark / GB10	25	14630	August 20, 2026
Qwen3.8-Flash-Next on 1, 2 and 4 DGX Sparks with NVIDIA's official NVFP4 quant: 64 tok/s peak single stream DGX Spark / GB10	27	5901	September 10, 2026
Qwen3.8-Flash-FP8 dual sparks DGX Spark / GB10	65	5045	September 7, 2026
Fastest Qwen 3.5 122B Int4 recipe on DGX Spark tested and published on Spark-Arena DGX Spark / GB10 llama	59	5491	June 3, 2026

Qwen 3.8 is about to launch, open weight too

Related topics