IDEAgent can operate entirely inside infrastructure you control. When configured exclusively with local models through Ollama, source code, retrieved context and model inference can remain inside the customer's own infrastructure.
Running serious coding models locally requires more than installing a model server.
Large models require GPU memory, PCIe bandwidth, host memory, storage, cooling and power engineered for sustained inference.
The IDEAgent Enterprise Node is designed as a complete local-AI platform rather than a collection of components.
Three planned configurations cover different deployment classes: 96 GB Team, 192 GB Ada Enterprise, and 384 GB Blackwell Enterprise.
Fully local AI is only as useful as the infrastructure underneath it.
There is a substantial difference between a machine that can technically load a model and a system capable of serving developers continuously without exhausting GPU memory, host memory, PCIe bandwidth or power capacity.
IDEAgent Enterprise Nodes are designed as complete systems: GPU memory for model residency, high-lane-count PCIe connectivity, ECC host memory, fast local storage, sustained cooling and redundant server power.
The goal is simple: the customer receives infrastructure designed to run IDEAgent locally instead of receiving another integration project.
IDEAgent's RAG and context infrastructure was developed on a deliberately modest workstation: a single NVIDIA GeForce RTX 3060 paired with an AMD Ryzen 5 2600X, running Ubuntu 24.04.4 LTS.
That machine hosts Ollama and the IDEAgent Docker stack and demonstrates an important point: IDEAgent itself does not require enterprise hardware.
Enterprise hardware becomes relevant when the requirement changes: larger models, higher precision, multiple simultaneous agents, larger teams and fully local inference.
| Team Node | Enterprise Ada | Enterprise Blackwell | |
|---|---|---|---|
| GPU | 2× RTX 6000 Ada | 4× RTX 6000 Ada | 4× RTX PRO 6000 Blackwell Max-Q |
| GPU memory | 96 GB ECC | 192 GB ECC | 384 GB ECC |
| Memory generation | GDDR6 | GDDR6 | GDDR7 |
| System RAM | 256 GB ECC | 512 GB ECC | 512 GB ECC |
| CPU | Threadripper PRO 7975WX | Threadripper PRO 7975WX | Threadripper PRO 7975WX |
| Primary purpose | Team inference | Large local models | Very large local models and higher concurrency |
| Indicative price | €32,175 ex-VAT | €53,287 ex-VAT | €86,614 ex-VAT* |
* Blackwell pricing is a current estimate. The final price will be confirmed after the complete 512 GB RAM and power-harness production configuration has been validated.
The premium configuration replaces the four 48 GB Ada GPUs with four NVIDIA RTX PRO 6000 Blackwell Max-Q GPUs.
Each GPU provides 96 GB ECC GDDR7, producing 384 GB of aggregate GPU memory across the four-GPU system.
The Max-Q variant is particularly interesting for this platform because each card retains a 300 W maximum power envelope while doubling memory capacity compared with the 48 GB Ada card.
* Current estimate. The Blackwell price will be confirmed once the production configuration is finalized, particularly the 512 GB ECC memory and the final GPU power-cabling solution.
| 192 GB Ada | 384 GB Blackwell | |
|---|---|---|
| GPUs | 4× RTX 6000 Ada | 4× RTX PRO 6000 Blackwell Max-Q |
| VRAM per GPU | 48 GB | 96 GB |
| Aggregate GPU memory | 192 GB | 384 GB |
| Host RAM | 512 GB ECC | 512 GB ECC |
| GPU memory generation | GDDR6 ECC | GDDR7 ECC |
| PCIe interface | PCIe Gen 4 | PCIe Gen 5 ×16 |
| GPU board power | 300 W each | 300 W each |
| Four-GPU board power | 1,200 W | 1,200 W |
| Primary advantage | Lower acquisition cost | Twice the GPU memory, substantially higher memory bandwidth and newer architecture |
Model parameter count alone does not tell you whether a model fits into GPU memory.
Precision matters.
A model stored in FP16 or BF16 requires approximately two bytes per parameter for the weights alone. Quantized models use significantly less memory.
Ollama commonly distributes large local models in quantized formats, so a statement such as “480B runs locally” must not be interpreted as “480B FP16 fits into VRAM.”
| Model class | Approx. native BF16 / FP16 weights | 192 GB Ada | 384 GB Blackwell |
|---|---|---|---|
| 32B | ~64 GB | Native | Native |
| 50B | ~100 GB | Native | Native |
| 70B | ~140 GB | Native | Native |
| 100B | ~200 GB | Quantization or offload | Native |
| 150B | ~300 GB | Quantization or offload | Native — workload dependent |
| 200B | ~400 GB | Quantization / offload | Quantization or limited offload |
| 400B+ | ~800 GB+ | Quantized / offloaded | Quantized / offloaded |
These figures represent weight memory only. Real inference also requires memory for the runtime, KV cache, context and concurrent requests.
Ollama provides downloadable model packages using specific quantization levels.
The package size is therefore more useful for practical deployment planning than parameter count by itself.
| Ollama model | Package / precision | Approx. size | 192 GB Ada | 384 GB Blackwell |
|---|---|---|---|---|
| Qwen3-Coder 30B | FP16 | 61 GB | Fully GPU resident | Fully GPU resident |
| Llama 4 Scout | Q4_K_M | 67 GB | Fully GPU resident | Fully GPU resident |
| Llama 4 Scout | Q8_0 | 117 GB | Fully GPU resident | Fully GPU resident |
| Llama 4 Scout | FP16 | 217 GB | CPU/GPU offload required | Fully GPU resident |
| Llama 4 Maverick | Q4_K_M | 245 GB | CPU/GPU offload required | Fully GPU resident |
| Qwen3-Coder 480B | Q4_K_M | 290 GB | CPU/GPU offload required | Fully GPU resident |
| DeepSeek V3.1 671B | Q4_K_M | 404 GB | Heavy CPU/GPU offload | Limited CPU/GPU offload |
| Llama 4 Maverick | Q8_0 | 428 GB | Heavy CPU/GPU offload | Limited CPU/GPU offload |
| Qwen3-Coder 480B | Q8_0 | 510 GB | Heavy CPU/GPU offload | CPU/GPU offload required |
| DeepSeek V3.1 671B | Q8_0 | 713 GB | Beyond combined nominal GPU + host capacity | Heavy CPU/GPU offload within nominal system capacity |
| Qwen3-Coder 480B | FP16 | 960 GB | Does not fit in installed GPU + host memory | Does not fit in installed GPU + host memory |
| DeepSeek V3.1 671B | FP16 | 1.3 TB | Does not fit | Does not fit |
“Fully GPU resident” refers to the model package fitting inside aggregate installed GPU memory. Runtime overhead, context and KV cache still require additional memory and may reduce practical context length or concurrency.
Multi-GPU residency also depends on the inference runtime being able to distribute the model across the installed GPUs.
Qwen3-Coder 480B provides a useful real-world example.
Ollama's local Q4_K_M package is approximately 290 GB.
| 192 GB Ada | 384 GB Blackwell | |
|---|---|---|
| Installed GPU memory | 192 GB | 384 GB |
| Qwen3-Coder 480B Q4_K_M | 290 GB | 290 GB |
| Minimum weight capacity outside GPU memory | ~98 GB | 0 GB |
| Result | CPU/GPU offload required | Model weights can reside entirely in GPU memory |
This is the primary reason the 384 GB configuration exists.
It does not merely offer more theoretical capacity: it moves current very-large quantized models from CPU-assisted deployment into the class of models whose weights can reside entirely across the GPU pool.
GPU memory and system memory serve different roles.
The Enterprise Node therefore does not describe GPU memory and host RAM as one shared VRAM pool.
| GPU memory | ECC host RAM | |
|---|---|---|
| 192 GB Ada Node | 192 GB | 512 GB |
| 384 GB Blackwell Node | 384 GB | 512 GB |
The 512 GB host-memory pool provides capacity for:
CPU offload increases model capacity but can reduce inference performance because host memory is substantially slower than local GPU memory and transferred data must cross the CPU/GPU interconnect.
For this reason IDEAgent will publish measured performance for offloaded model configurations instead of predicting performance from memory capacity alone.
Enterprise deployment is not necessarily about assigning every available GPU to one enormous model.
Additional GPU memory can instead support multiple simultaneous IDEAgent workloads:
For many organizations, concurrency may be more valuable than maximum model size.
The 384 GB Blackwell configuration therefore provides two different advantages: very-large model residency and substantially more room for concurrent local workloads.
Large context windows consume additional GPU memory through KV cache.
IDEAgent's context architecture is designed specifically to avoid treating the model context window as repository storage.
Instead of repeatedly sending an entire repository to a model, the Context Engine retrieves relevant symbols, code, project knowledge, documentation and previous changes for the current task.
This allows model context and KV-cache allocation to be tuned around the actual coding workload, leaving more GPU memory available for model weights and concurrent inference.
The Ada and Blackwell Enterprise configurations use the same underlying server architecture.
| Component | Enterprise specification | Purpose |
|---|---|---|
| CPU |
AMD Ryzen Threadripper PRO 7975WX
32 cores / 64 threads |
High-lane-count platform for multi-GPU inference and supporting services. |
| Motherboard | ASUS Pro WS WRX90E-SAGE SE | Seven physical PCIe 5.0 ×16 slots, multi-GPU support and eight-channel ECC R-DIMM memory. |
| System memory | 512 GB DDR5 Registered ECC | Ollama runtime, model loading, CPU offload, RAG, embeddings and containers. |
| Storage |
2× 4 TB Samsung 990 Pro NVMe
RAID 1 |
Fast mirrored storage for models, indexes, containers and IDEAgent data. |
| Chassis | Supermicro SuperChassis 747BTQ-R2K04B | Rack-capable high-airflow chassis selected for dense multi-GPU operation. |
| Power | 2× 2000 W redundant Titanium hot-swap | 1+1 redundant power sized around the four-GPU configuration. |
| Cooling | High-static-pressure chassis cooling + active TR5 CPU cooling | Sustained airflow for dense local inference workloads. |
The RTX PRO 6000 Blackwell Max-Q is designed for dense workstation installations and supports configurations from one to four GPUs.
Each GPU provides 96 GB ECC GDDR7 memory, uses a dual-slot form factor, active cooling, PCIe Gen 5 ×16 and a maximum 300 W power envelope.
This makes the card particularly well matched to the IDEAgent Enterprise platform: four cards provide 384 GB aggregate GPU memory while retaining the same nominal 1,200 W four-GPU power envelope as the existing Ada design.
The production configuration will nevertheless undergo complete physical validation before shipment.
Validation includes:
The Enterprise power architecture is designed around maximum component load rather than average inference consumption.
| Component | Quantity | Peak per unit | Combined estimate |
|---|---|---|---|
| RTX 6000 Ada or RTX PRO 6000 Blackwell Max-Q | 4 | 300 W | 1,200 W |
| AMD Threadripper PRO 7975WX | 1 | 350 W | 350 W |
| WRX90 motherboard + 512 GB ECC RAM | 1 system | Estimated | ~120 W |
| NVMe, fans and supporting hardware | 1 system | Estimated | ~100 W |
| Estimated maximum component load | — | — | ~1,770 W |
Production systems will be power and thermal tested under sustained multi-GPU load before these engineering estimates are replaced with measured values.
| Configuration | Hardware | Indicative selling price |
|---|---|---|
| 96 GB Team | 2× RTX 6000 Ada | €32,175 ex-VAT |
| 192 GB Enterprise | 4× RTX 6000 Ada | €53,287 ex-VAT |
| 384 GB Blackwell Enterprise | 4× RTX PRO 6000 Blackwell Max-Q | €86,614 ex-VAT* |
* Current estimate. The Blackwell price will be confirmed once the production configuration is finalized, particularly the 512 GB ECC memory and the final GPU power-cabling solution.
For organizations considering dedicated private GPU infrastructure, an on-premises system changes the economics from recurring GPU rental to a customer-owned capital asset with local electricity, cooling and support costs.
Cloud GPU pricing varies substantially by provider, accelerator, region, commitment level and utilization.
IDEAgent therefore does not claim one universal cloud hourly rate.
A meaningful TCO analysis should use the customer's real workload and compare it with the complete lifecycle cost of the node.
| Cost area | Cloud GPU infrastructure | IDEAgent Enterprise Node |
|---|---|---|
| Infrastructure | Recurring GPU-instance rental | Customer-owned hardware |
| Compute usage | Usually scales with runtime | No external GPU-hour charge for local inference |
| Electricity | Included in provider pricing | Customer electricity and cooling |
| Infrastructure location | Selected cloud environment | Customer-controlled infrastructure |
| Asset after lifecycle | No physical GPU asset | Hardware remains customer property |
IDEAgent does not need the largest model available to work.
The hardware exists to give organizations a choice: smaller efficient local models, large native-precision models, very-large quantized models, or several simultaneous agent workloads.
Designed for teams running smaller and medium local models, RAG and agent workloads without requiring a four-GPU system.
The cost-focused Enterprise configuration.
It provides enough GPU memory for large native models in roughly the 70B class, substantial quantized models, multiple smaller models and CPU-assisted deployment of larger MoE models.
The high-capacity Enterprise configuration.
Its 384 GB GPU pool moves models such as Qwen3-Coder 480B Q4_K_M into the fully GPU-resident class while retaining 512 GB of ECC host memory for IDEAgent, runtime services and larger-model offload.
It also provides substantially more capacity for concurrent developers and parallel agents.
The product is therefore not tied to one model.
It is a local inference platform whose GPU capacity, host memory and software architecture can continue to serve IDEAgent as the model ecosystem changes.
The final product family is:
Final production specifications, measured inference performance and pricing will be published after physical system validation.
IDE