The Best AI Model for a Small Cluster Is the One That Fits

Last Updated: 16 September 2026

Which open-weight model should run on a 256GB dual-machine AI cluster? The one whose weights actually fit, not the one topping the leaderboards. DeepSeek's new V4.1 Flash is faster and cheaper through its API, but GLM 5.3 Flash is the model with a documented, working recipe serving on two NVIDIA DGX Spark units today. On small hardware, fit decides the winner before benchmarks get a vote.

What DeepSeek actually shipped on 10 September

DeepSeek V4.1 Flash is a 552-billion-parameter open-weight model with a one-million-token context window, released on 10 September under an MIT licence. Its design activates only 8 billion parameters per input token and 16 billion per output token, cutting its cache memory to roughly a quarter of the previous generation, and its API prices start at $0.15 per million input tokens off-peak. It is, on paper and in hosted benchmarks, the new value benchmark for open-weight AI.

DeepSeek V4.1 Flash, released on 10 September, is a 552-billion-parameter mixture-of-experts model with a one-million-token context window and support for up to 384,000 output tokens, according to DeepSeek's own release. Its distinguishing design is a causal encoder-decoder that activates only 8 billion parameters per token when reading input and 16 billion when generating, which is aimed squarely at input-heavy agent workloads. The model card also reports a global KV cache of roughly 890 bytes per token, about a quarter of its predecessor's, with the persistent cache footprint cut to roughly an eighth. Weights are published under the MIT licence.

The pricing moved too. DeepSeek's published rate card now lists uncached input at $0.15 per million tokens off-peak and $0.30 at peak, with output at $0.60 and $1.20 respectively. Off-peak rates are half of peak, so scheduling batch work outside weekday morning windows halves the token bill. Two days after release, DeepSeek began routing requests for its flagship V4-Pro name to V4.1 Flash at Flash rates, a strong signal about which model the company now considers its workhorse.

What the independent benchmarks say

Independent measurement puts DeepSeek V4.1 Flash at roughly twice the hosted output speed of GLM 5.3 Flash, 198 tokens per second against 90.5. But GLM holds the most recent published quality score, 57 on the Artificial Analysis intelligence index against 52 for DeepSeek's previous Flash, and V4.1 Flash has not been scored yet. Speed and quality currently point at different models, which is precisely why the fit question in the next section matters.

According to Artificial Analysis, which measures competing models through an identical harness, DeepSeek V4.1 Flash generates 198 tokens per second on its hosted API against 90.5 for GLM 5.3 Flash, with a faster time to first token as well. On raw throughput and responsiveness, the newer DeepSeek model is roughly twice as fast as GLM's offering. Speed is not the whole story, though: on Artificial Analysis's intelligence index, GLM 5.3 Flash scored 57 against 52 for DeepSeek's previous V4 Flash, and no intelligence score for V4.1 Flash had been published at the time of writing.

Read the table as two different arguments: DeepSeek wins today's speed measurement, GLM holds the last published quality score. Neither settles which model belongs on your own hardware.

The catch: what fits on a desk-sized cluster

Hosted benchmark results say nothing about whether a model's weights fit the hardware a small business could actually own. A two-unit NVIDIA DGX Spark cluster pools 256GB of unified memory, and GLM 5.3 Flash is the model with a published, working recipe on exactly that hardware, serving at 21.8 tokens per second. No equivalent recipe for DeepSeek V4.1 Flash exists yet, so today the fit winner is GLM even though the speed winner is DeepSeek.

The benchmark table above measures hosted APIs, where someone else's data centre absorbs the weight of the model. Running the weights yourself inverts the question. Each NVIDIA DGX Spark unit pairs 128GB of unified memory with a GB10 chip, so a two-unit cluster pools 256GB, and that number, not the leaderboard, sets which models are candidates at all.

The GLM recipe that works

Within days of GLM 5.3 Flash's release, a developer published a full deployment recipe on NVIDIA's developer forums for serving the model across two DGX Spark units with tensor parallelism. The checkpoint, a 182-GiB quantised build, fits the pooled memory; the documented result was 21.8 tokens per second decode with a 262,144-token context window, measured as medians of three runs of 200-token greedy streaming. That is not fast by data-centre standards, but it is a complete, reproducible proof that this class of model runs on desk-sized, affordable hardware.

Where DeepSeek V4.1 Flash stands

No comparable documented two-unit recipe for V4.1 Flash existed at the time of writing. The model is days old, and its published deployment guidance targets data-centre class hardware. One developer on the same NVIDIA forums reported decode speeds around 350 tokens per second through the hosted API, which underlines the point: the fastest way to use V4.1 Flash today is to rent it, not to run it. That is a statement about maturity and fit, not about the model's quality.

What this means when you buy AI capacity

For most small-business workloads, renting models like these through an API is cheaper than owning hardware, and the choice between them is decided by fit: whether the model runs on the hardware you have, serves your workload's context needs, and costs what the rate card says. Buy a documented, measured serving recipe over a leaderboard position, and rent capability unless data locality or cost structure genuinely demands owning the machines.

According to the pricing DeepSeek has published, a million input tokens processed off-peak costs fifteen cents, a price at which the hosted API is almost certainly cheaper than owning any hardware for a typical small-business workload. The case for local hardware rests on other grounds: data that must not leave the building, flat costs at high volume, and independence from API changes such as the silent rerouting of V4-Pro requests that DeepSeek began on 14 September.

If local capacity is genuinely the requirement, the lesson of this fortnight is to buy fit first and capability second. A documented serving recipe with published measurements, like the GLM one above, is worth more than a benchmark table entry for a model that has no path onto your hardware yet.

In our own discovery conversations, the hardware question comes up early, and owners often assume that using AI means buying machines. Our delivery model takes the opposite position: a small self-hosted server orchestrating agent work, with scoped API access to models like these, captures most of the value at a fraction of the capital. The fit-versus-benchmarks question in this story is exactly the analysis we run per workload, and most workloads resolve to renting.

Sources

• Artificial Analysis, DeepSeek V4.1 Flash vs GLM 5.3 Flash comparison: https://artificialanalysis.ai/models/comparisons/deepseek-v4-1-flash-vs-glm-5-3-flash
• Artificial Analysis, GLM 5.3 Flash vs DeepSeek V4 Flash comparison: https://artificialanalysis.ai/models/comparisons/glm-5-3-flash-vs-deepseek-v4-flash
• DeepSeek, "Introducing DeepSeek-V4.1-Flash" announcement: https://www.deepseek.com/en/news/deepseek-v4-1-flash/
• DeepSeek API Docs release notes, 10 September 2026: https://api-docs.deepseek.com/news/news260910
• DeepSeek-V4.1-Flash model card, Hugging Face: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
• NVIDIA Developer Forums, GLM 5.3 Flash on 2x DGX Spark deployment recipe: https://forums.developer.nvidia.com/t/glm-5-3-flash-on-2x-nvidia-dgx-spark-43-4-tok-s-peak-checkpoint/381429
• NVIDIA Developer Forums, DeepSeek V4.1 Flash on DGX Spark thread: https://forums.developer.nvidia.com/t/deepseek-v4-1-flash/382725