Vibe coding workstation: four NVIDIA RTX PRO Blackwell kits from 32 GB to two 96 GB cards

Which workstation for vibe coding with a local model

For one person who builds apps by describing them to a local model, the smallest workstation that works is one NVIDIA RTX PRO 4500 Blackwell with 32 GB: by our users-per-card rule it holds Qwen3.8-27B at 4-bit with one agent session of 128K tokens. An RTX PRO 5000 with 48 or 72 GB runs the same model at 8-bit with longer or more sessions, one RTX PRO 6000 Max-Q or Workstation Edition with 96 GB runs it in 16-bit or a larger model, and two of them serve a small studio. Beyond that, the model belongs on a server.

Vibe coding here means building an app, an internal tool or a prototype by telling an AI tool what you want, checking the result and correcting it in words. The tool, such as Dyad, OpenHands or a terminal agent, sends the model long prompts with instructions, files and error output on every step. Our comparison of self-hosted Lovable alternatives covers the tools, and our article on whether a local model is good enough for vibe coding covers quality against Claude. This guide turns both into four kits. The general GPU comparison for every kind of AI work is in our guide to the best workstation GPU for AI.

What sets the size: weights, context and people

Three things decide how much GPU memory a vibe coding workstation needs.

The weights. A model's parameters are stored at 16, 8 or about 4 bits each. Qwen3.8-27B, a dense Apache 2.0 model that Qwen published in August 2026, is 56 GB in Ollama's 16-bit (BF16) file, 30 GB in its 8-bit q8_0 file and 18 GB in its 4-bit q4_K_M file.

The context. Every open session keeps a key-value cache, the memory where the model holds everything it has read in that session. Coding tools fill it quickly: Ollama asks for "at least 64000 tokens" of context for them, the OpenHands guide sets 22,000 as the minimum, and Qwen's card for Qwen3.6-27B recommends keeping at least 128K tokens to preserve the model's thinking capability. Qwen3.8-27B keeps a full cache in 16 of its 64 layers, with 4 key-value heads of dimension 256: by our arithmetic from its config.json that is 64 KiB per token in 16-bit, 8 GiB for 128K tokens, plus about 0.14 GiB of fixed state per session for its linear-attention layers.

The people. One maker who is working holds one session. A second builder window, a background agent or a colleague adds another. Speed is shared between the sessions that generate at the same moment.

Ollama library tags and Hugging Face file lists, read on 30 October 2026, converted to GiB; Qwen3-Coder-Next and gpt-oss from our coding-assistant guide. Cache per 128K-token session in 16-bit by our arithmetic from each config.json, including fixed state; an FP8 cache in vLLM halves the per-token part.

The memory a card offers comes from our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card for the CUDA context and activations, as our guide to how many users one RTX PRO 6000 serves explains. That gives 25.7 GiB on a 32 GB card, 40.0 GiB on 48 GB, 61.5 GiB on 72 GB and 83.0 GiB on 96 GB.

Four kits from 32 GB to two 96 GB cards

The table maps each kit to the model it runs and the number of 128K sessions that fit beside it.

Our estimates by the users-per-card rule with the weights and Ollama's default 16-bit cache from the table above; not measurements. RTX PRO 4500 and 5000 driver memory scaled from the 95.6 GiB of a 96 GB card. The RTX PRO 6000 Workstation Edition (600 W) has the same 96 GB and gives the same counts as the Max-Q (300 W).

Kit 1, RTX PRO 4500 Blackwell (32 GB, 896 GB/s, 200 W). Qwen3.8-27B at 4-bit leaves 8.9 GiB: one session of 128K or two of 64K. For quick layout changes, Qwen3.6-35B-A3B, a mixture-of-experts (MoE) model that computes about 3B of its 35B parameters per token, leaves 3.3 GiB at 4-bit, two sessions of 64K, and answers far faster. GLM-4.7-Flash fits too, at "~23 GB VRAM" for 64,000 tokens by Ollama's figure. The 8-bit file of the 27B model (27.9 GiB) does not fit, and neither does a 256K context. The 24 GB RTX PRO 4000 below it keeps 18.5 GiB by our rule, too little for a 64K session beside a 27B model at 4-bit; it runs gpt-oss-20b with three sessions of 64K, as our page on 30B models on 24 GB cards explains for that class. Our page to buy an RTX PRO 4500 workstation lists the multi-card builds.

Kit 2, RTX PRO 5000 Blackwell (48 or 72 GB, 1,344 GB/s, 300 W). On 48 GB, Qwen3.8-27B at 8-bit leaves 12.1 GiB, one session of 128K or two of 64K, and Qwen3.6-35B-A3B in FP8 leaves room for three of 64K. On 72 GB the 8-bit model keeps four sessions of 128K or two of 256K, and even the BF16 file fits with one 128K session. gpt-oss-120b leaves only about 0.7 GiB on 72 GB, short of one 64K session. Our comparison of the RTX PRO 5000 with 48 or 72 GB covers the rest of the difference.

Kit 3, one RTX PRO 6000 Blackwell Max-Q or Workstation Edition (96 GB, 1,792 GB/s). The Workstation Edition draws up to 600 W, the Max-Q 300 W, with the same memory and bandwidth. Qwen3.8-27B in BF16 keeps three sessions of 128K; Qwen's FP8 checkpoint (28.7 GiB) in vLLM keeps six, or 13 with an FP8 cache; at 32K that is 47 sessions, the count our other Qwen3.8-27B pages use. This is also the kit for an 80B-class MoE model: Qwen3-Coder-Next in FP8 holds five sessions of 64K, and its card states that it supports only non-thinking mode, so it starts writing code without a reasoning phase. gpt-oss-120b holds nine sessions of 64K.

Kit 4, two RTX PRO 6000 Max-Q or Workstation Edition (192 GB). The simplest layout runs one copy of Qwen3.8-27B per card, six sessions of 128K each. One model can also span both cards as a tensor-parallel pair in vLLM: Qwen3-Coder-Next in FP8 then holds about 57 sessions of 64K by our rule, as our coding-assistant guide calculates. The cards have no NVLink, so a pair exchanges data over PCIe, as our article on the RTX PRO 6000 without NVLink explains. Our guide to what a workstation with two Max-Q runs covers other models.

For each of these kits you can order the workstation alone, the setup of Ollama, the model and the context alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Published speeds for one maker

One measurement of these models on RTX PRO Blackwell cards has been published: StorageReview tested Qwen3.8-27B in NVFP4, in a build the review does not name, with vLLM 0.27.1 (NGC container 26.08) in a review published on 14 October 2026. At one request with 512-token prompts and answers, every card landed between 17 and 22 tokens per second. With 32 requests at once, an RTX PRO 6000 Workstation Edition in a Dell Pro Max Tower T2 delivered 616 tokens per second in total at 600 W, and 425 with 8K-token prompts and 1K answers; the RTX PRO 4500 Blackwell Server Edition, the 165 W server card with 800 GB/s, delivered 320 and 153. We found no published result for the workstation RTX PRO 4500 or the RTX PRO 5000 with these models.

For your own estimate, the ceiling for one request is the memory bandwidth divided by the bytes read per token: the weights less the embedding table, of which each token reads only one row. The 8-bit model reads about 28 GB per token, which gives about 63 tokens per second on an RTX PRO 6000 Workstation Edition or Max-Q and about 47 on an RTX PRO 5000; the 18 GB 4-bit file gives about 50 or a little more on an RTX PRO 4500. These are upper bounds by our estimate (about 56 tokens per second on the Server Edition in a rack server), and real speeds are lower, as the 17 to 22 tokens per second StorageReview measured for one request show. A MoE model with 3B active parameters reads far less per token and is several times faster. Thinking adds time as well: Qwen3.8-27B reasons at its highest effort by default, and the model card offers medium and low settings. Our article on speculative decoding explains how multi-token prediction (MTP) and draft models raise single-request speed.

What more memory buys in quality

More memory buys a better model before it buys more precision. In Qwen's own table for Qwen3.6, the dense 27B scores 77.2 on SWE-bench Verified against 73.4 for the 35B-A3B MoE model, and Qwen3.8-27B scores 61.7 on SWE-bench Pro against 53.5 for Qwen3.6-27B in the Qwen3.8 table. The step from 4-bit to 16-bit is smaller: NVIDIA's card for its NVFP4 build of Qwen3.8-27B lists 74.02 on Terminal-Bench against 75.56 for BF16. That figure is for NVIDIA's NVFP4 format; the q4_K_M files in Ollama use a different 4-bit method, for which we found no comparable published figure. These are vendor figures, and our article on local models against Claude for vibe coding sets them in context. The practical order for spending memory: first a 27B-class model with 64K to 128K of context, then 8-bit or 16-bit weights, then more sessions.

Software on the kit

Ollama GitHub releases and blog of 23 January 2026, Hugging Face model cards and the projects' GitHub pages, read on 30 October 2026; Cline and Continue from our coding-assistant guide, OpenCode and Codex CLI from our guide to Claude Code with a local model.

Ollama 0.35.0 or later matters because earlier versions are affected by CVE-2026-103663, as our coding-assistant guide records. The context length is set in Ollama's settings or with OLLAMA_CONTEXT_LENGTH; Ollama's documentation ties the default to GPU memory, 4K below 24 GiB, 32K up to 48 GiB and 256K above, so set the planned 64K or 128K explicitly. Ollama also answers one request per model at a time unless OLLAMA_NUM_PARALLEL is raised, and each parallel slot reserves the full context; the session counts above assume both settings. ollama launch, available since Ollama v0.15, starts Claude Code, OpenCode, Codex or Droid against a local model, and our guide to Claude Code with a local model covers that path. Claude Code is proprietary and used under Anthropic's terms, so it is not part of the kit; its user installs it. bolt.diy is left out for a similar reason: the WebContainers runtime it builds on needs a commercial licence for production use in a for-profit company, as our comparison of self-hosted Lovable alternatives explains. Ollama binds to 127.0.0.1 by default; on a kit that others reach over the network, the model server goes behind access control, as our guide to securing a GPU server describes. Before an app built on the kit goes live, our security checks for vibe-coded apps list what to test.

Tell us which tools your team already uses, and our team replies with a configuration for your tasks and budget.

Desk or rack: power, noise and Max-Q

By our estimate, with a 350 W-class processor and 15 per cent for memory, drives, fans and power-supply losses, a kit with one RTX PRO 4500 draws about 0.6 kW at the wall, one RTX PRO 5000 about 0.75 kW, one RTX PRO 6000 Workstation Edition about 1.1 kW, two Max-Q about 1.1 kW and two Workstation Editions about 1.8 kW. Each fits a 230 V 16 A circuit planned at 80 per cent, 2.94 kW, if nothing large shares the breaker; a measurement on the delivered system confirms the figure.

NVIDIA states up to four Max-Q cards in one workstation. Puget Systems found in July 2025 that directly adjacent Max-Q cards have little room to draw air and recommends a slot gap; Exxact reports, in an article updated in March 2026, that all four adjacent cards throttled without its custom airflow and held 300 W at up to 85 °C with it. Three or four Max-Q therefore belong only in a chassis built for them, with strong directed airflow. Two RTX PRO 6000 Workstation Editions, 600 W each with flow-through coolers, are the realistic maximum for that edition, as our guide to two Workstation Editions, heat and power shows. Our comparison of the three RTX PRO 6000 editions sets out their coolers and power side by side. Whether a desk, a side room or a rack suits your office is the subject of our guide to rack server, tower or workstation for AI.

When a workstation stops and a server starts

A workstation suits one to a handful of makers. When a development team of 20 or more works with agents all day, the model moves to a rack server with RTX PRO 6000 Server Edition cards, the edition we recommend for continuous operation in a rack, as our guide to a server for coding agents sizes for 20 to 100 developers. Models of the 1T class need a server too: Kimi K2.7-Code is 595.2 GB in INT4 and runs on eight RTX PRO 6000 Server Edition cards, as our guide to Kimi K2.7-Code on eight Server Edition cards shows.

Give us the number of people who build and the tools they use: describe your plan in a few lines, and we propose the kit for your tasks and budget.

FAQ