DeepSeek V4.1 Flash
2.00 bpw · 257 GBMeasured- GPUGeForce RTX 5090 D V2 24GB
- RAM128 GB DDR5-4800
- CPUIntel Core i7-14700KF
Price alert
We'll email you when the Amazon.jp price reaches your target.
Released in China in August 2025, the RTX 5090 D V2 drops to 24GB VRAM but still runs Qwen3.8 27B at 4-bit at about 57 tokens/s, faster than a 4090. Listings often leave out the V2, so go by the 24GB of VRAM.
Some buy links are affiliate links: we may earn a commission if you buy through them, at no extra cost to you.
A single 5090 D V2 reaches 35B-class at 4-bit, with 30B-class MoE at about 150 tokens/s. Context is tight: Gemma 4 31B gets only about 60K.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 3.8 GB | Runs | 193.7 | 128K | |
| 4.1 GB | Runs | 181.6 | 256K | |
| 4.7 GB | Runs | 161.7 | 256K | |
| 10.5 GB | Runs | 155.9 | 256K | |
| 11.8 GB | Runs | 76.3 | 192K | |
| 9.8 GB | Runs | 89.5 | 221K | |
| 11.9 GB | Runs | 164.9 | 198K | |
| 11.8 GB | Runs | 73.7 | 143K | |
| 12.3 GB | Runs | 173.1 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 5 GB | Runs | 158 | 128K | |
| 5.7 GB | Runs | 141.6 | 256K | |
| 7.1 GB | Runs | 117.4 | 256K | |
| 16.9 GB | Runs | 136 | 256K | |
| 16.8 GB | Runs | 55.7 | 117K | |
| 16.5 GB | Runs | 56.6 | 122K | |
| 18.3 GB | Runs | 149.5 | 115K | |
| 18.3 GB | Runs | 50.3 | 66K | |
| 22.1 GB | Runs | 152.2 | 123K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 8.2 GB | Runs | 106 | 128K | |
| 9.5 GB | Runs | 93 | 256K | |
| 12.7 GB | Runs | 71.7 | 256K | |
| 26.9 GB | Needs RAM offload | 27.1 | 256K | |
| 28.6 GB | Too big | — | — | |
| 29 GB | Too big | — | — | |
| 31.8 GB | Needs RAM offload | 23.1 | 198K | |
| 32.6 GB | Too big | — | — | |
| 36.9 GB | Needs RAM offload | 35.8 | 256K |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Needs RAM offloadMoE only: expert weights sit in system RAM, so speed is estimated from 70 GB/s RAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
With Strata, GeForce RTX 5090 D V2 can run large models that far exceed its VRAM capacity—the bulk of the model stays in RAM and SSD, while the GPU holds only the most frequently used experts. Benchmarks come from the Strata author and community, with pricing filled in by VRAMGlass at current market rates.
Some buy links are affiliate links: we may earn a commission if you buy through them, at no extra cost to you.
The same 27B to 31B dense model runs more than 30% faster, with Qwen3.8 27B at about 57 tokens/s
Qwen3.6 35B A3B and GLM 4.7 Flash at 4-bit both run at around 150 tokens/s without touching system RAM
24GB is just enough, and Qwen3.8 27B can run up to 122K of context
That takes 29 GB, more than 24GB, so this tier calls for the original 32GB 5090 D
Its compute is about 10% lower than even the 4090's
It takes 575W to get 24GB, while the 4090 draws 450W and the 3090 350W
Both versions are called RTX 5090 D: the V2 has 24GB, the original 32GB. Listing titles often leave out “V2”, so check the VRAM capacity in GPU-Z or nvidia-smi; if it shows 24GB, it is this version.
NVIDIA's China website lists the V2's length and slot width as “varies by manufacturer”, so check the specific model before you build. Before launch, Benchlife reported that manufacturers can raise the power limit of factory-overclocked versions to 600W.
As with the 5090, the official recommendation is a 1000W power supply. The connector is 12V-2x6, and with the adapter all four 8-pin plugs must be connected. This connector has melted on the 5090; for the causes and how to prevent it, see the RTX 5090 page.
When the original 5090 D launched, there were claims that its driver restricts multi-GPU and AI workloads, which NVIDIA did not confirm. We haven't found reliable tests showing whether the V2 behaves the same way, so confirm before buying several.
GeForce RTX 5090 D V2 vs GeForce RTX 4090, GeForce RTX 5090 D, and GeForce RTX 3090: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | ||||
|---|---|---|---|---|
| Memory | 24 GB | 24 GBSame | 32 GB33% more | 24 GBSame |
| Bandwidth | 1,344 GB/s | 1,008 GB/s25% lower | 1,792 GB/s33% higher | 936 GB/s30% lower |
| BF16 / FP16 compute | 297 TFLOPS | 330 TFLOPS11% higher | 297 TFLOPSSame | 142 TFLOPS52% lower |
| INT8 compute | 594 TOPS | 661 TOPS11% higher | 594 TOPSSame | 285 TOPS52% lower |
| Qwen3.8 27B speed (2-bit) | 89.5 t/s | 69.3 t/s23% slower | 114.6 t/s28% faster | 64.8 t/s28% slower |
| Largest dense model (4-bit) | 37B | 37BSame | 51B38% more | 37BSame |
| Power | 575 W | 450 W22% lower | 575 WSame | 350 W39% lower |
| GPU interconnect | Not supported | Not supported | Not supported | NVLink 112.5 GB/s |
| Interface | PCIe 5.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x16 | PCIe 4.0 x16 |
| Released | 2025-08 | 2022-10 | 2025-01 | 2020-09 |
| Amazon.jp median |
Some buy links are affiliate links: we may earn a commission if you buy through them, at no extra cost to you.
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.