Research notes for Dell Technologies
What two Dell Pro Max with GB10 units can run, how fast they run it, and how far behind the best cloud models they are.
Since the units arrived in late July we have run open-weight language models on two of these boxes, around the clock. They power our agents, our research tools and the experiments behind this report. This is what we found, written for the people at Dell who build them.
The short version: a model that matches the frontier of 9 July now runs on two units on a desk. It reads 840,000 tokens in one prompt, and no third-party model API is involved.
We measure the gap in time, not points. For each model that runs on the Dell units, we find the day the frontier, the highest score any model had reached on the index, first matched it, then count to the day the open model was released.
Two things keep this honest. The index tests each model as its developer serves it, at full precision and maximum reasoning effort; our runs use compressed weights. And the frontier did not wait: it stood at 42 for fifteen days before Claude Opus 5 reached 51 on 24 July.
The gap is weeks, not years. And it runs on hardware you can carry.
Each unit has 128 GB of memory shared by the CPU and GPU, about 121.7 GiB of it usable. The model's weights, its working memory for your conversation (the KV cache) and the operating system all have to fit. When they don't, a second unit joins over the ConnectX-7 link and each box holds half the model.
The GB10 pairs 128 GB of memory with 273 GB/s of bandwidth, shared by CPU and GPU. Every token it writes means streaming the model's active weights through that pipe. So on this machine, fewer bits per weight does two things at once: a bigger model fits, and each token has fewer bytes to read.
That makes the choice of compression format the most important decision on a GB10, and not all 4-bit formats are equal. In an independent test of GLM-5.3-Flash, the EXL3 4-bit build we run drifted from the original model exactly as much as the official FP8 release did on the same software, at 54% of the size. NVIDIA's native NVFP4 format, at about the same size, drifted 2.5 times as much.
This test is someone else's, and our deployment adds FP8 dense layers it did not cover. So our next report measures it directly: the same public evaluation questions, asked of our 4-bit build on the Dell pair and of Z.ai's full-precision GLM API.
GLM-5.3-Flash is the model we live with. These numbers are ours, from acceptance tests on 8 September 2026 on the two Dell units, with 850,000 tokens of context enabled and thinking turned off.
Single-stream output speed, tok/s, median of three runs
One prompt, nothing cached, read at 1,011 tok/s
Asked about details planted at the start, middle and end, it answered all three correctly.
To first token on a repeated 47k-token agent prompt. 76 s the first time, before the prefix was cached.
Total output with four requests in flight at the full 850k context setting.
Prompt reading speed at 8k tokens, cold. 1,238 tok/s at 32k.
What would a full day of the pair's work cost in the cloud? Assume it runs 24 hours at our measured maximum. The fair comparison is the closed model with the same score, GPT-5.6 Terra, at its list price.
Agents read far more than they write: every turn re-sends the conversation, tool results and documents. 20 is our example; drag to match your workload. The pair runs 24 hours a day at its measured maximum.
Daily capacity uses our measured rates on the pair: 1,238 tok/s reading cold prompts and 59.8 tok/s writing across four requests, one after the other. Cached prompts read far faster, so real capacity is higher. Cloud prices at list, with no caching discounts. Electricity is left out: each GPU drew 40 to 75 W in our tests.
Three lessons from running the pair around the clock. Each one would make anyone's first week with these boxes easier.
On these units host RAM is GPU memory, so the operating system, containers and the model share one 128 GB pool. The tightest moment is a long prompt, not boot. We size the KV cache explicitly (14 GiB for GLM at 850k context) rather than by a utilization percentage, and leave the host room to breathe.
The ConnectX-7 link carries the tensor-parallel traffic, and the recipes can share the head unit's 164 GiB checkpoint to the worker over it instead of copying. When one unit's internet link was faster, we downloaded there and moved the files over the cable.
An 840k-token prompt takes about 14 minutes before the first token. Proxies and CDNs time out long before that, so we added keepalive messages to the stream. Anyone serving long context from these boxes will hit the same wall.
Six open recipes, each exposing an OpenAI-compatible endpoint. Our own numbers are marked. The rest are reported by Mia's AI Lab on NVIDIA DGX Spark, which uses the same GB10 chip and memory; re-running them on our Dell pair is next.
From Artificial Analysis model pages, checked 24 September 2026, at each model's highest effort setting: GPT-5.4 (39, 5 March), Claude Opus 4.7 (41, 16 April), GPT-5.6 Terra (42, 9 July), Claude Opus 5 (51, 24 July), Claude Fable 5.1 (53, 1 September), Claude Opus 5.5 (58, 22 September). Local models: Qwen3.8-Flash-Next and GLM-5.3-Flash (26 August), DeepSeek V4.1 Flash (10 September).
Measured with sparkDash against a running server, thinking off. Decode rates exclude time to first token; cold prefill uses unique prompts. Each figure names its configuration; numbers from different configurations are not mixed in one row. Our units: Dell Pro Max with GB10 FCM1253, BIOS 5.36, DGX OS 7.4.0, driver 580.173.02, CUDA 13.0, linked by a 200 Gb/s ConnectX-7 cable.
Index scores are for full-precision models; local builds are compressed (GLM 4-bit, which matched official FP8 on the same software in an independent KL-divergence test; DeepSeek 2.9-bit, not yet quality-checked). Field Report 02, planned for October 2026, compares our local build against Z.ai's full-precision API on the same questions.