Free tools Windows power users keep installed
One-click scans. No signup required.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Yes, a single desktop RTX 5090 can run Qwen3.8-27B, but there is no one reliable tokens-per-second figure for every setup. In Tom’s Hardware’s September 8, 2026 tests, a single-card SGLang configuration reached 37,740 tokens of context; that did not mean it could also handle the model’s full 262,144-token window with the desired prompt-processing time and throughput. Separate measurements report faster decode rates under different quantization, software, and prompt conditions, so compare complete setups rather than headline speeds.
How fast is Qwen3.8-27B on an RTX 5090?
It depends on the exact workload and software stack. 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. The repository identifies its setup as vLLM 0.27.1 in WSL2 and says the measurements came from its own machine: Murai Labs’ repository.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
ASUS TUF Gaming GeForce RTX™ 5080 16GB GDDR7 OC Edition Graphics Card | $1,830.91 | Buy on Amazon |
Those figures are useful as a concrete example, not a universal RTX 5090 benchmark. They are not a controlled comparison with Tom’s Hardware’s tests, which used different configurations and evaluated factors beyond decode speed. Prompt length matters: a speed measured with a short prompt does not establish how quickly the same setup will process a long prompt or respond at a large context.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Can one RTX 5090 run the model’s full context?
Tom’s Hardware’s September 8, 2026 testing found that a single RTX 5090 could run Qwen3.8-27B for some shorter-context workloads. In its tested SGLang configuration, the reported context reached 37,740 tokens and was substantially faster than the publication’s single-card vLLM recipe. The result is specific to the tested versions and configuration; it does not establish a general ranking of SGLang and vLLM.
#1 Best Overall
- Powered by the NVIDIA Blackwell architecture and DLSS 4. System Requirements: Minimum 850W PSU with 16-pin 12V-2x6 (12VHPWR) connector required. Verify before purchasing.
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability. Compatibility: 348mm (13.7") length, 3.6 slots, 4.3 lbs. Confirm case clearance and slot spacing. GPU bracket included.
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.6-slot design with massive fin array optimized for airflow from three Axial-tech fans
- Phase-change GPU thermal pad helps ensure optimal thermal performance and longevity, outlasting traditional thermal paste for graphics cards under heavy loads
The same article says its single-card setup could not combine the model’s full native 262,144-token context with reasonable prompt-processing times and high throughput. That distinction is important: fitting a model and some context is not the same as achieving a useful response time at the maximum context. For that combined goal, Tom’s Hardware says to start above 32 GB of VRAM or use multiple cards. Read its benchmark and configuration details at Tom’s Hardware.
What changes the benchmark result?
A GPU’s model name or VRAM capacity alone does not predict usable performance. Compare the details that determine how much memory remains for context and how quickly the chosen engine processes and generates tokens.
- Quantization and KV cache: Reduced-precision model weights and cache formats affect memory use and may involve fidelity tradeoffs. A result using NVFP4 weights or NVFP4 KV cache is not directly interchangeable with one using another format.
- Inference engine and software: Engine version, kernels, patches, and operating environment can change both feasibility and speed. Results from an ordinary install should not be assumed to match a patched stack.
- Prompt and context length: Decode speed at a short prompt says little about prompt-processing time or performance as context grows. Check the actual prompt length and context limit reported for the test.
- Measurement method and concurrency: Time-to-first-token, prompt-processing speed, decode throughput, and the number of concurrent requests describe different parts of serving performance. A tokens-per-second figure without its workload and measurement conditions is incomplete.
- Task type: Image or video input and long-context agent workflows can impose different requirements from short text conversations. Use results that reflect the task you intend to run.
What do two-card and older-GPU results show?
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 window. Its setup uses vLLM 0.29.0rc2, a patched software stack, and NVFP4 KV cache; the project records a measurement session dated September 23, 2026. This demonstrates that a specifically engineered two-card configuration can target the full window, not that an unpatched system or any two-card setup will reproduce it. See the project configuration.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →RTX 4090 and RTX 3090
In the llama.cpp tests described by Tom’s Hardware, the RTX 4090 and RTX 3090 used Q8_0 KV cache, with about 112K tokens as the practical context limit in those configurations before VRAM ran out. The author also reports an unexplained long-context performance cliff on the RTX 4090 and says it may indicate a bug. Treat that observation as a limitation of the reported test, not a settled conclusion about every RTX 4090 setup.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to choose a benchmark that matches your workload
- Set the context target first. Decide whether you need short conversations, long documents, or the full 262,144-token window. Do not use a maximum-window result as a proxy for speed at ordinary prompts.
- Check memory available after loading. Account for model weights, runtime allocations, and KV cache, rather than comparing the GPU’s advertised VRAM with model-weight size alone.
- Match the formats and engine. Compare quantization, cache format, inference engine and version, kernels, patches, and operating environment. A change in any of these can make a benchmark a poor match.
- Compare the right timing measures. Look for prompt-processing time or time-to-first-token alongside decode speed, and note prompt length and concurrency. Decide which matters most for your use: an interactive single request or serving multiple requests.
- Verify the task. If your use includes image or video inputs or long-context agent work, prioritize benchmarks that reflect those demands rather than a short text-generation test.
What 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: Qwen’s model card. NVIDIA’s NGC listing documents BF16 and FP8 versions, SGLang, Linux, and the hardware it tested. NVIDIA also says it did not independently reproduce the evaluation results linked from that listing: NVIDIA NGC’s listing. These sources describe the model and a vendor’s tested software and hardware; they do not replace consumer RTX 5090 measurements under a reader’s own workload.
If you are considering an RTX 5090 for local inference, the card alone does not guarantee the speed or context capacity in any cited benchmark. Choose based on the context, prompt-processing, and throughput targets you actually need, then match the software and memory configuration to those targets.
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.
