> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tokios.com/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> Tokios exposes private local models through https://api.tokios.com using an outbound-only connector tunnel.
> Use 'connector' for the local tokios-connector process.
> Use 'deployment' for the public model name in the API model field.
> Use 'upstream model id' for the model served by Ollama, llama.cpp, vLLM, LM Studio, or another local backend.
> API keys use the prefix sk-tok-. Connector tokens use the prefix ct-tok-.

# RTX PRO Blackwell variants guide

> Compare RTX PRO 4000, 4500, 5000, and 6000 Blackwell professional GPUs. Workstation vs Max-Q vs Server vs 'D' variants — which fits your use case.

NVIDIA's RTX PRO Blackwell lineup spans 4 VRAM tiers (24 GB to 96 GB) and 3 form factors (Workstation, Max-Q, Server), plus a China export-compliance variant. This guide explains the differences so you pick the right SKU for your workload.

> [!NOTE]
>
> Last verified: 2026-10-21. Prices and availability change; confirm with NVIDIA or your distributor before buying.

## The RTX PRO Blackwell lineup

| Model | Memory | Bandwidth | TDP | Form factor | Price (MSRP) | Best for |
| - | - | - | - | - | - | - |
| **RTX PRO 4000** | 24 GB GDDR7 | 672 GB/s | 150W | Workstation | \~\$2,000-2,500 | 30B Q4, 14B FP16 |
| **RTX PRO 4500** | 32 GB GDDR7 | 896 GB/s | 200W | Workstation | \~\$3,000-4,000 | 70B Q4, 32B dense FP16 |
| **RTX PRO 5000** | 48 GB or 72 GB GDDR7 | 1,344 GB/s | 250W | Workstation | \~\$5,000-7,000 | 70B FP8, 120B MoE Q4 |
| **RTX PRO 6000 Workstation** | 96 GB GDDR7 | 1,792 GB/s | 300W | Workstation (full tower) | \~\$16,000 | 120B MoE FP8, 70B FP16 |
| **RTX PRO 6000 Max-Q** | 96 GB GDDR7 | 1,597 GB/s | 250W | Low-profile (4x per chassis) | \~\$8,500 | 4-GPU workstation (384 GB total) |
| **RTX PRO 6000 Server** | 96 GB GDDR7 | 1,792 GB/s | 300W | Rack-mount (datacenter) | varies | High-density serving |
| **RTX 6000D** | 84 GB GDDR7 | 1,597 GB/s | 300W | Workstation | varies | China market (export-compliance) |

> [!TIP]
>
> **The "D" suffix** on RTX 6000D means "China domestic" — it's an export-compliance variant with 84 GB instead of 96 GB (\~12% lower performance). Available only through authorized China distributors.

## Form factor decision tree

### Workstation (full tower)

**RTX PRO 6000 Workstation** is the flagship: 96 GB GDDR7 in a standard full-tower chassis with active cooling and display outputs.

**When to choose:**

* You need maximum single-GPU VRAM (96 GB) for 120B MoE models at FP8
* You have a standard workstation tower with adequate cooling
* You need display outputs for interactive work
* Budget is \~\$16,000 per card

**Example use cases:**

* Serving DeepSeek V4-Flash-0731 (162-167 GB) on 2x PRO 6000 Workstation (192 GB total)
* Running GLM-4.5-Air (106B MoE) at FP8 on a single card
* Multi-user team serving with 70B models at high precision

### Max-Q (low-profile, 4x per chassis)

**RTX PRO 6000 Max-Q** is the density play: same 96 GB GDDR7 as the Workstation variant, but in a low-profile form factor that fits 4 cards in a single workstation chassis.

**When to choose:**

* You need 384 GB total VRAM in one machine (4x 96 GB)
* You're running frontier models like DeepSeek V4.1 Flash (763B total params) at high precision
* You have a workstation chassis designed for 4x low-profile GPUs
* Budget is \~\$8,500 per card (cheaper than Workstation variant)

**Example use cases:**

* 4x RTX PRO 6000 Max-Q (384 GB total) runs DeepSeek V4.1 Flash at FP8+MXFP4 with \~135 tok/s per GPU
* High-density inference server for multiple 70B models
* Research workloads that need 300+ GB VRAM without clustering multiple nodes

> [!WARNING]
>
> **Max-Q requires the right chassis.** Not all workstation cases fit 4x low-profile GPUs. Check PCIe slot spacing, power delivery (4x 250W = 1,000W total), and cooling airflow before ordering.

### Server (rack-mount, datacenter)

**RTX PRO 6000 Server** is identical to the Workstation variant in specs, but designed for rack-mount chassis with datacenter cooling (no active fans on the card, relies on chassis airflow).

**When to choose:**

* You're building a rack-mount inference server
* Your datacenter provides forced-air cooling
* You need high-density deployment (multiple cards per 2U/4U chassis)
* You don't need display outputs

**Example use cases:**

* Multi-tenant inference API serving
* Enterprise LLM deployments with SLA requirements
* Clusters of 8-16 cards for large-scale model serving

### RTX 6000D (China export-compliance)

**RTX 6000D** is a cut-down RTX PRO 6000 for the China market. It has 84 GB GDDR7 instead of 96 GB, with \~12% lower bandwidth and performance.

**When to choose:**

* You're in China or selling to China-based customers
* You need to comply with U.S. export controls on advanced AI hardware
* 84 GB is sufficient for your workload (still fits 120B MoE at Q4)

**Example use cases:**

* Same as RTX PRO 6000 Workstation, but with 12 GB less VRAM
* 2x RTX 6000D (168 GB total) runs DeepSeek V4-Flash-0731 with less headroom than 2x PRO 6000

## RTX PRO 4000 / 4500 / 5000: Mid-tier professional cards

Not every workload needs 96 GB. The mid-tier RTX PRO cards offer professional features (ECC memory, ISV certifications, professional drivers) at lower price points.

### RTX PRO 4000 (24 GB, \~\$2,000-2,500)

**Best for:** 30B models at Q4, 14B models at FP16

**Example workloads:**

* Qwen3.6-35B-A3B (35B MoE, 3B active) at Q4: \~75 tok/s
* Llama 3.1 14B at FP16: \~75 tok/s
* Gemma 4 E4B at Q8: \~100 tok/s

**When to choose over RTX 4090:**

* You need ECC memory for production serving
* You need ISV certifications for professional software
* You need professional drivers with longer support lifecycle
* You need 24/7 reliability guarantees

### RTX PRO 4500 (32 GB, \~\$3,000-4,000)

**Best for:** 70B models at Q4, 32B dense models at FP16

**Example workloads:**

* Qwen3 70B at Q4: \~85 tok/s
* Llama 3.1 32B at FP16: \~85 tok/s
* 2x RTX PRO 4500 (64 GB total) runs 120B MoE models with tensor parallelism

**When to choose over RTX 5090:**

* Same 32 GB VRAM as RTX 5090, but with professional features
* \~15% slower than RTX 5090 on raw inference speed, but with ECC and ISV support
* Better for production deployments where reliability matters more than peak speed

### RTX PRO 5000 (48 GB or 72 GB, \~\$5,000-7,000)

**Best for:** 70B models at FP8, 120B MoE models at Q4

**Example workloads:**

* Qwen3 70B at FP8: \~100 tok/s (higher quality than Q4)
* Qwen3.5-122B-A10B (122B MoE) at Q4: \~30 tok/s
* GLM-4.5-Air (106B MoE) at Q4: \~35 tok/s

**When to choose:**

* You need 48-72 GB VRAM for 70B models at FP8 (better quality than Q4)
* You're running 120B MoE models that don't fit in 32 GB
* You need professional features but can't afford 96 GB PRO 6000

**48 GB vs 72 GB variant:**

* 48 GB: Fits 70B FP8 with \~8 GB headroom for KV cache
* 72 GB: Fits 70B FP8 with \~32 GB headroom, or 120B MoE at higher quantization

## Multi-GPU configurations

RTX PRO cards shine in multi-GPU setups. Here are common configurations:

### 2x RTX PRO 6000 (192 GB total)

**Best for:** DeepSeek V4-Flash-0731 (162-167 GB), GLM-5.2 at Q4

**Performance:**

* DeepSeek V4-Flash-0731 at FP8+MXFP4: \~106-210 tok/s (community-reported range)
* GLM-5.2 at Q4: \~26 tok/s (estimated, needs validation)

**Requirements:**

* Workstation chassis with 2x full-length PCIe 5.0 x16 slots
* 2x 300W PCIe power connectors (or 12VHPWR)
* NVLink bridge recommended for tensor parallelism (RTX PRO 6000 supports NVLink)

### 4x RTX PRO 6000 Max-Q (384 GB total)

**Best for:** DeepSeek V4.1 Flash (763B total params) at high precision

**Performance:**

* DeepSeek V4.1 Flash at FP8 dense + MXFP4 experts: \~135 tok/s per GPU (540 tok/s aggregate)
* Multiple 70B models simultaneously

**Requirements:**

* Workstation chassis designed for 4x low-profile GPUs
* 4x 250W PCIe power connectors (1,000W total)
* High-airflow cooling (4x cards generate significant heat)

### 2x RTX PRO 4500 (64 GB total)

**Best for:** 120B MoE models at Q4, cost-effective alternative to 2x PRO 6000

**Performance:**

* Qwen3.5-122B-A10B at Q4: \~40-50 tok/s (estimated)
* GLM-4.5-Air at Q4: \~45 tok/s (estimated)

**Requirements:**

* Workstation chassis with 2x full-length PCIe 5.0 x16 slots
* 2x 200W PCIe power connectors
* NVLink bridge recommended (RTX PRO 4500 supports NVLink)

## Professional features vs consumer cards

Why pay 2-3x more for RTX PRO over RTX 5090?

| Feature | RTX PRO Blackwell | RTX 5090 | Why it matters |
| - | - | - | - |
| **ECC memory** | Yes (optional) | No | Prevents silent data corruption in production |
| **ISV certifications** | Yes | No | Guaranteed compatibility with professional software |
| **Professional drivers** | Yes (longer lifecycle) | Yes (shorter lifecycle) | Stability over bleeding-edge features |
| **NVLink support** | Yes (all models) | No | Fast multi-GPU interconnect (900 GB/s) |
| **24/7 reliability** | Yes (rated for continuous operation) | No (rated for 8h/day) | Critical for production serving |
| **Warranty** | 3 years (standard) | 2 years (standard) | Longer coverage for enterprise deployments |
| **Form factors** | Workstation, Max-Q, Server | Workstation only | Density options for datacenter |

> [!TIP]
>
> **When to choose RTX PRO:**
>
> * Production serving with SLA requirements
> * Multi-tenant API serving where downtime costs money
> * Enterprise deployments with compliance requirements
> * Multi-GPU configurations that need NVLink
>
> **When to choose RTX 5090:**
>
> * Individual developer or small team
> * Experimentation and research (not production)
> * Budget-constrained projects
> * Single-GPU workloads that don't need NVLink

## Pricing and availability

| Model | MSRP (USD) | Availability | Notes |
| - | - | - | - |
| RTX PRO 4000 | \~\$2,000-2,500 | Q4 2026 | Entry professional |
| RTX PRO 4500 | \~\$3,000-4,000 | Q4 2026 | Mid-tier professional |
| RTX PRO 5000 (48 GB) | \~\$5,000-6,000 | Q4 2026 | 48 GB variant |
| RTX PRO 5000 (72 GB) | \~\$6,000-7,000 | Q4 2026 | 72 GB variant |
| RTX PRO 6000 Workstation | \~\$16,000 | Q4 2026 | Flagship workstation |
| RTX PRO 6000 Max-Q | \~\$8,500 | Q4 2026 | Density-optimized |
| RTX PRO 6000 Server | varies | Q4 2026 | Contact NVIDIA sales |
| RTX 6000D | varies | Q4 2026 | China only |

> [!WARNING]
>
> **Availability is limited at launch.** RTX PRO Blackwell cards are produced in lower volumes than consumer RTX cards. Lead times may be 4-8 weeks. Order through authorized NVIDIA professional partners, not consumer retail channels.

## Runtime compatibility

All RTX PRO Blackwell cards have full CUDA support and work with all major inference backends:

| Backend | Support | Notes |
| - | - | - |
| **Ollama** | Yes | Simplest setup, single-user |
| **vLLM** | Yes | Production-grade, concurrent batching |
| **SGLang** | Yes | High-throughput, RadixAttention |
| **llama.cpp** | Yes (CUDA) | CPU+GPU offloading, GGUF quantization |
| **LM Studio** | Yes | Desktop GUI, easy model management |
| **TensorFold** | Yes | EXL3 quantization, highest throughput |

> [!TIP]
>
> **NVLink matters for multi-GPU.** If you're running 2x or 4x RTX PRO cards, use NVLink for tensor parallelism. PCIe 5.0 x16 is 128 GB/s per direction; NVLink is 900 GB/s. The 7x bandwidth difference translates to 10-20% higher throughput on large models.

## Next steps

> **[Hardware reference](/guides/hardware-reference.md)**
>
> Compare RTX PRO against consumer GPUs, Mac Studio, DGX Spark, and budget hacker tier.

> **[DGX Spark model compatibility](/guides/dgx-spark-model-compatibility.md)**
>
> See which models fit on DGX Spark's 128 GB unified memory.

> **[Model sharding](/guides/model-sharding.md)**
>
> Configure tensor parallelism across multiple GPUs.

> **[DeepSeek V4.1 Flash on 2 DGX Sparks](/blog/deepseek-v41-flash-dgx-spark.md)**
>
> Community benchmarks for running V4.1 Flash on 2 DGX Sparks (alternative to 2x RTX PRO 6000).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.