Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesYes. A desktop RTX 5090 can run Qwen3.8-27B, but there is no single reliable tokens-per-second figure for every setup. Published results vary with quantization, inference software, prompt length, context size, and measurement method. Tests reported in 2026 show that a single card can be useful for shorter-context workloads; running the model’s full 262,144-token window with fast prompt processing and high throughput is a different, more demanding target.
How fast is Qwen3.8-27B on an RTX 5090?
It depends on what is being measured. Decode speed after the prompt has been processed is not the same as prompt-processing speed or time to first token. A result measured with a short prompt also does not predict performance after a long conversation has filled much of the context window.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
ASUS TUF Gaming GeForce RTX 5090 32GB GDDR7 OC Edition Gaming Graphics Card | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
Murai Labs reports 107.6 tokens per second at a prompt length of 256 and 97.3 tokens per second at a prompt length of 2,048 for NVFP4 Qwen3.8-27B on one 32 GB desktop RTX 5090. Its repository identifies the setup as vLLM 0.27.1 in WSL2 and says the measurements came from its own machine. These are useful configuration-specific reference points, not a guaranteed speed for other systems.
Tom’s Hardware’s September 8, 2026 testing offers a separate comparison of inference setups. In its reported single-card tests, SGLang ran substantially faster than its single-card vLLM recipe and reached a context of 37,740 tokens. That result does not establish a universal SGLang advantage: software versions, kernels, settings, and workload all matter, and its result is not directly comparable to Murai Labs’ figures.
#1 Best Overall
- AI Performance: 772 AI TOPS
- OC mode: 2580 MHz Default mode: 2550 MHz(Boost clock)
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- SFF-Ready Enthusiast GeForce Card
- Axial-tech fans feature a smaller fan hub that facilitates longer blades and a barrier ring that increases downward air pressure
Why benchmark numbers differ
- Prompt length: Processing input can take substantial time before generated tokens begin. A speed figure without prompt length and time-to-first-token leaves out important parts of the experience.
- Context and KV cache: Longer contexts consume more memory for the key-value cache. Cache format and context settings affect both how much fits and how quickly it runs.
- Quantization: NVFP4, FP8, BF16, and other formats have different memory and computation requirements. A speed result for one format does not transfer automatically to another; quantization also involves a fidelity trade-off.
- Inference stack: Engine and version, kernels, patches, and runtime allocations influence performance and memory use.
- Workload: Concurrency, generation length, and whether the task includes image or video input change what a meaningful benchmark should measure.
Can one RTX 5090 handle the full 262K context?
Not with the combination of full native context, prompt processing, and throughput demonstrated in Tom’s Hardware’s tested single-card configurations. The article’s single-card SGLang setup reached 37,740 tokens of context, while its tests did not achieve the full 262,144-token window with the desired prompt-processing times and throughput. For that full-context target, the article points to starting above 32 GB of VRAM or using multiple cards.
That is a workload-specific finding, not a claim that the model cannot load or generate on a single RTX 5090. Shorter contexts can be practical even when the complete model window is not. The relevant question is whether your application needs the entire window at once, and whether it also needs responsive prompt ingestion and sustained generation speed.
What changes with two RTX 5090 cards?
A public project by Adrien Brault documents a Qwen3.8-27B NVFP4 serving configuration spanning two RTX 5090 cards and a 262,144-token context. Its measurement session is dated September 23, 2026. The setup uses vLLM 0.29.0rc2, a patched software stack, and NVFP4 KV cache.
This shows that a two-card configuration can be built for the full context under specified conditions. It does not show that two cards with default software settings will produce the same result, nor does it provide a directly comparable speed result to the single-card reports. Multiple GPUs add complexity as well as capacity: software support, interconnect and system configuration, memory allocation, and the chosen cache format all matter.
What about RTX 4090 and RTX 3090 cards?
Tom’s Hardware reports that its RTX 4090 and RTX 3090 llama.cpp tests used Q8_0 KV cache and treated about 112K tokens as the practical context limit in the described configurations, based on what fit before running out of VRAM. The article also notes an unexplained long-context performance cliff on the RTX 4090 and says it may indicate a bug. That anomaly should not be treated as a normal expected speed result.
These observations are specific to those tests and their software and cache settings. They are not a clean ranking of the RTX 4090, RTX 3090, and RTX 5090: the setups and feasible contexts differ, and an older card’s behavior can change with engine, quantization, and cache choices.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to judge a benchmark for your workload
Before using a published tokens-per-second result to choose hardware or estimate responsiveness, check whether it matches the task you intend to run.
- Match the model format. Check whether the benchmark uses NVFP4, FP8, BF16, or another quantization, and account for the fidelity implications as well as memory use.
- Check usable VRAM, not just card capacity. Model weights are only one allocation; runtime overhead and the KV cache also need memory.
- Match context and cache settings. Compare the actual prompt and context lengths, and note the KV-cache format. A short-context result is not evidence of full-window performance.
- Identify the software stack. Record the inference engine, version, relevant patches, and platform. A patched configuration should not be assumed to work like an unmodified installation.
- Look beyond decode speed. Check prompt-processing time or time to first token as well as generation speed, and note concurrency and the benchmark’s measurement method.
- Match the input type. If your use case includes images or video, confirm that the tested model and serving path cover that workload rather than relying on a text-only speed number.
What the official model and hardware listings establish
Qwen’s model card describes Qwen3.8-27B as a multimodal model for images and videos and includes serving examples. That information helps establish model capabilities and ways to serve it; it is not a consumer RTX 5090 speed test.
NVIDIA’s NGC listing documents BF16 and FP8 versions, SGLang, Linux, and the hardware NVIDIA lists as tested. The listing also says NVIDIA did not independently reproduce the linked model evaluation results. Consumer RTX 5090 performance figures therefore need to be read in the context of the separate editorial and community setups that measured them.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
