Running GLM-5.3-Flash on CMP 170HX — a sizing guide

---

Read this before the numbers. Every tokens-per-second figure, every RAM minimum and every "fits / does not fit" on this page is calculated from memory bandwidth, not measured. We publish it because the arithmetic is useful for choosing parts, and because nobody else seems to have written it down. We will replace it with real figures once we have run the model — the benchmark is planned on an 8-channel, 512 GB machine and has not happened yet. Plan for the lower end of every range.

GLM-5.3-Flash runs on one or two unlocked CMP 170HX 64 GB cards when the rest of the model sits in system RAM. With two cards and 128 GB of RAM it should run at good quality; the amount and speed of system memory decide how fast it answers. From four cards on, the good-quality model fits entirely on the cards and system RAM stops mattering.

What we have actually measured

These are ours, from our own bench and from a four-card system we commissioned in September 2026. They are the foundation the estimates are built on.

Everything below this line is modelling.

How it works

GLM-5.3-Flash is a mixture-of-experts model: 320 billion parameters in total, but only 18 billion are used for each word. The card holds the parts every word needs; the rest stays in system RAM, and only the few experts each word picks are read from there.

Those parameter counts and the 288-experts-per-layer layout are Z.ai's published figures, read from the model card and its config.json. They are not ours to verify.

Pick a build

Two cards with 128 GB of RAM is the build to aim for: good answer quality at a usable speed. The other two are a cheaper start and a single-card option.

A token is roughly three quarters of a word. 10 tokens/s reads about as fast as a person reads; 20+ feels quick.

How much system RAM you need

RAM needed = the part of the model that does not fit on the cards + about 16 GB for the operating system. One card leaves about 46 GB of its memory for model weights; two cards leave about 105 GB.

The smallest version (93 GB) does not fit on one card alone, so a single-card system always needs a large RAM pool. File sizes are unsloth's published GGUF sizes, not ours.

Expected speed

The number of memory channels matters more than the CPU model. A second card helps most on desktop platforms, because it moves a bigger share of the model out of slow RAM.

More cards: 3 to 6

From three cards on, the whole model fits in card memory, so system RAM no longer limits speed.

• More cards buy quality and room, not speed per word. The cards work one after another, and a higher-quality file reads more data for each word, so speed falls slightly from four cards to six.
• Four cards is the sweet spot for one user: good quality, about 30 tokens/s, no fast-RAM platform needed.
• Six cards suit a team. The official FP8 weights fit and a vLLM server could answer several users at once. Space for context is tight at about 37 GB, and vLLM support for this model still needs checking.
• What changes from three cards: a workstation or server board with one ×16 slot per card (EPYC or Threadripper; ×8 slots also work, only loading gets slower), a power supply of about 250 W per card plus 300 W (about 1.3 kW for four, 1.8–2 kW for six), and a chassis with airflow for every card.

These estimates assume each card keeps about 58 GB for model weights and reaches about 35% of its 950 GB/s on this model — typical for similar models on cards of this generation, but an assumption, not a result.

System checklist

The rows that most often go wrong are memory channels, Above 4G decoding and cooling.

On the PCIe link: we have measured the link state (Gen2 ×16). We have not benchmarked transfer throughput on it. A Gen2 ×16 link tops out around 8 GB/s in theory and real traffic lands nearer 6–7 GB/s after overhead — treat that as derived from the specification, not as a measurement.

What to expect

• Long prompts start slowly. Reading a long document or a large code base takes seconds to minutes before the first word appears, longer with one card. The Gen2 link is the limit there; answer speed is not affected.
• Lower-bit versions lose quality. UD-Q2_K_XL works but makes more mistakes than UD-Q4_K_XL. For coding or reasoning, choose the 4-bit version.
• Long context is cheap. Most layers keep a fixed-size state, so a long conversation costs little extra card memory.
• New model, young software support. GLM-5.3-Flash uses a new attention design. llama.cpp loads it, but parts may run less efficiently than they could; this improves with updates.
• The 64 GB comes from a patched driver. After a kernel update the cards fall back to 8 GB until the unlock is re-installed.
• No ECC. The card does not correct or report memory errors, so use cards that have passed a full memory test. Ours ship with a per-card memory verification report carrying that card's serial number.

Where the numbers come from

Sources: zai-org/GLM-5.3-Flash · unsloth/GLM-5.3-Flash-GGUF

---

commissioned here. Every CMP 170HX we sell is memory-tested at full 64 GB capacity and ships with a report carrying its own serial number. If you want this build sized against your actual workload, write to us.*