Open-weight models like Kimi K3 and Qwen3.8 Max reach frontier level. But between “weights available” and sensible local operation lie several terabytes of VRAM, high-performance networking and serious power requirements.
Matthias Allitsch-Wutte · GEKI founder
4 August 2026 · 8 min read
We wanted to know what Kimi K3 and Qwen3.8 Max really mean on your own hardware. So we did the math.
The result up front: open-weight models reach frontier level - but between “weights available” and sensible local operation lie several terabytes of VRAM, high-performance networking and double-digit kilowatts. And for most companies the right answer isn’t the biggest model, but the smallest one that solves their task reliably.
Because at GEKI we plan and operate exactly this kind of infrastructure, this isn’t an armchair estimate: it’s the same sizing we do for real deployments.
Kimi K3 and Qwen3.8 Max compete with leading proprietary models on coding, reasoning and agentic tasks.
Both use a Mixture-of-Experts (MoE) architecture. The active compute is therefore closer to a model of around 100 billion parameters. Yet the memory footprint is set by all the trillions of parameters. That is where the real hardware problem comes from.
VRAM is the especially fast memory directly on the GPUs. It holds the model weights, temporary computations and the KV cache for context and parallel users. If the model doesn’t fit entirely into this fast memory, data has to move between GPU, RAM or SSD. That can make a model technically startable, but far too slow for interactive production use.
Rule of thumb for the raw weights:
On top of that comes memory for scaling values, non-quantized model parts, context, framework and communication buffers. Newer model generations often no longer train or ship full 16-bit weights, but native 4-bit. That cuts the memory footprint a lot: a model with 2.8 trillion parameters needs around 1.4 TB at 4-bit. The released Kimi K3 checkpoint (MXFP4) sits at roughly 1.56 TB due to additional model components - without native 4-bit weights it would be far larger.
Kimi K3 activates only 104 billion of 2.8 trillion parameters per token, Qwen3.8 Max about 95 billion of 2.4 trillion. A router decides for each token which experts are used. That’s why all expert weights have to be immediately available.
So sparse MoE mainly reduces the compute per token. It does not reduce the required memory in the same proportion. 95 billion active parameters sound manageable, but still need terabytes of VRAM.
Three models as examples for different size tiers: Kimi K3 (frontier), DeepSeek V4 Flash (mid-tier) and Qwen3.8 27B (compact). For each model, two setups: Minimal for a local experiment and Production for inference with concurrent users.
Mid-tier and compact models can run efficiently on local hardware with NVIDIA RTX PRO 6000 (96 GB) and AMD Instinct MI350P (144 GB) - up to 8× in a single server. At frontier size that is no longer enough: you need a data center with cards like NVIDIA H200 (141 GB, previous generation) or B200 (192 GB, current generation).
Weights = model size. KV cache = free VRAM after the weights (context, concurrent users, runtime).
| Model | Weights / KV cache | Hardware | Requirements |
|---|---|---|---|
| Kimi K3 · Minimal | 1.56 TB / ~0.70 TB | 16× H200 (2.26 TB) | Data center, air cooling possible, ~11 kW |
| Kimi K3 · Production | 1.56 TB / ~1.5 TB | 16× B200 (3.1 TB) | Modern high-density data center, DLC, ~16 kW |
| DeepSeek V4 Flash · Minimal | ~160 GB / ~32 GB | 2× RTX PRO 6000 (192 GB) | Single server, ~1.2 kW |
| DeepSeek V4 Flash · Production | ~160 GB / ~224 GB | 4× RTX PRO 6000 (384 GB) | Single server, ~2.4 kW |
| Qwen3.8 27B · Minimal | ~31 GB / ~65 GB | 1× RTX PRO 6000 (96 GB) | Local workstation, ~0.6 kW |
| Qwen3.8 27B · Production | ~31 GB / ~161 GB | 2× RTX PRO 6000 (192 GB) | Single server, ~1.2 kW |
For scale: Kimi K3 in production means 16× B200 - typically two servers with 8 GPUs each. The GPUs alone draw ~16 kW; with power supplies, CPUs, storage, networking and cooling, system draw sits clearly higher, often roughly 20+ kW. As monthly hardware financing (60 months), the two GPU servers sit at about €9,000 each, so ~€18,000/month - plus at least ~€10,000/month for power and operations of the two machines, before staff and software.
By comparison: at GEKI we run DeepSeek V4 Flash as mid-tier today on 2× or 4× RTX PRO 6000 in a regional data center - one server, around 1.2–2.4 kW, without high-density DLC and without that monthly cost base. Once the MI350P servers arrive, we will run the same workload on 2× MI350P for even better price-performance.
The weights are free. The infrastructure is not. Running a frontier model in production means:
One model, two fully loaded B200 servers and 20+ kW of system power: that is no longer a single GPU server, but a data-center project in its own right.
Frontier models like Kimi K3 or Qwen3.8 Max pay off at the top end: autonomous coding agents across very large repositories, hard scientific or technical reasoning, long-running agents with many tools and little oversight, or cases where the last few points of model quality carry high economic value.
Mid-tier workhorse models like Qwen3.8 27B or DeepSeek V4 Flash, by contrast, cover the bulk of real enterprise work: internal knowledge assistants and RAG, document analysis, support agents, coding with human oversight, domain agents for finance, legal or IT, and privacy-critical applications in your own data center.
The core point: most companies don’t need the smartest available model, but the smallest one that solves their concrete task reliably. And that line keeps moving: better training data, distillation, quantization and more efficient inference fit more capability onto the same server every generation.
Open frontier AI on your own infrastructure is possible, but expensive: Kimi K3 in production needs 16× B200, two servers, high-performance networking, DLC and roughly 20+ kW of system power - monthly about ~€18,000 hardware financing plus ~€10,000 and more for power and operations.
The mid-tier class below changes the hardware requirements fundamentally. DeepSeek V4 Flash (284 billion parameters, 13 billion active) we run at GEKI today on 2× or 4× RTX PRO 6000 in a regional data center. Once MI350P is available, we will run 2× MI350P for even better price-performance. In vendor benchmarks (not independently reproduced) it scores 82.7 on Terminal Bench 2.1, ahead of a strong proprietary mid-tier model like GPT-5.6 Terra (78.4).
Instead of 16 data-center GPUs and five-figure monthly costs, two to four GPUs in a single server are enough: a fraction of the accelerators, power and cost - a little less peak performance, which is simply enough for the vast majority of workloads. That’s exactly the class we build and operate at GEKI, on NVIDIA RTX PRO 6000 and soon AMD Instinct MI350P.
From a single server to larger configurations - GEKI plans, delivers and operates it with SLAs.
