Can I Run GLM-5.3-Flash Locally? Stock llama.cpp v0.6.0, Not a 24 GB Card

Short answer: Yes, on stock llama.cpp v0.6.0 (tag published 2026-10-05), if the machine can hold the weights. That tag’s notes add GLM-5.3-Flash (GLM5-Next). The GGUFs ggml-org published on 2026-10-03 are Q2_K at 124.62 GiB and Q4_K at 168.92 GiB. A 24 GB card is the wrong machine. A 128 GB unified box is not a measured resident fit for those official files. Unsloth’s smaller dynamic quants are a separate download, and their docs still print a fork clone. LocalRig has not measured tok/s on this model. Mac numbers now exist from other people, with labels: DwarfStar’s SSD-streaming measurement, an MLX-Serve vendor claim, and one llamaperf community report. The mechanism page is SSD expert streaming on a Mac.

This page is a constraint selector. It is not a quality ranking against Claude or GPT, and it is not the older GLM-5.2 flagship guide. Mixing those two names is how people buy a 24 GB card for a 100 GB-class weight file.

Is GLM-5.3-Flash the same model as GLM-5.2?

No. GLM-5.2 is Zhipu’s larger 1M-context flagship; the LocalRig 5.2 page still treats it as a datacenter / 256 GB-class problem. GLM-5.3-Flash is the smaller Flash SKU. Unsloth documents it as 320B parameters, 18B active. The ggml-org model card says the same 320B / 18B active figure, plus 288 routed experts and 1 shared expert. The vLLM recipe (file date_updated: 2026-09-18) says approximately 321B total and 18B active. Those are two publisher counts. LocalRig does not average them into a third size.

The GGUF header on ggml-org/GLM-5.3-Flash-GGUF (Q2_K first shard, read 2026-10-06) stores general.architecture = glm5-next, general.name = GLM-5.3-Flash, and glm5-next.context_length = 1,048,576. The same header’s general.size_label is 288x10B. That string matches the expert count in the card. It is not a second parameter count, and it is not 2.88 trillion weights.

Unsloth’s quality table compares Flash against GLM-5.2 and several closed models. That is a vendor scorecard, not a LocalRig ranking, and it is not a VRAM table.

Can stock llama.cpp run GLM-5.3-Flash now?

The v0.6.0 notes say yes. LocalRig has not loaded the file. llama.cpp v0.6.0 was published 2026-10-05. The notes list “GLM-5.3-Flash (GLM5-Next): 320B KDA/DSA hybrid text+vision model with mHC and MoE” and point at pull #27773. That pull merged 2026-09-30. The v0.6.0 changelog is “since v0.5.0” and is the first tagged release whose notes include this model. A build from a tag older than v0.6.0 is the wrong binary.

Two earlier llama.cpp pulls for this architecture closed without merging on 2026-10-01: #27754 and #27752. Do not check those branches out. Unsloth’s Hugging Face card still links #27754. That link is stale.

MTP for this model was a different pull, #27917. Checked again 2026-10-21, that pull is still open. A later pull, #29928, says it supersedes #27917, merged 2026-10-07, and llama.cpp v0.7.0 (published 2026-10-15) lists “GLM5-Next: MTP speculative decoding” under that number. The v0.6.0 notes do not include it. Speculative decoding on stock v0.6.0 for this SKU is not yet measured. LocalRig has not run the v0.7.0 MTP path either.

Which files exist, and how large are they?

Two official quants, plus Unsloth’s dynamic set. Sizes below are byte sums from the Hugging Face tree API on 2026-10-06, not a guess and not a single ~150 GB file.

ggml-org (upstream quantizer)

Repo ggml-org/GLM-5.3-Flash-GGUF, created 2026-10-03, last modified the same day. Base model on the card: zai-org/GLM-5.3-Flash-BF16. The card’s own notes: Q4_K puts routed experts at Q4_K; Q2_K puts routed down experts at Q4_K and gate/up experts at Q2_K; small non-expert tensors stay Q8_0; no imatrix on the low-bit quants. Label: upstream / ggml-org, not a LocalRig quant.

File	Bytes (tree API)	GiB	What it is
GLM-5.3-Flash-Q2_K (2 shards)	133,809,056,384	124.62	Weights. Shard 1 is 9,429,952 bytes (header). Shard 2 holds the tensors.
GLM-5.3-Flash-Q4_K (2 shards)	181,372,463,744	168.92	Weights, same split shape.
mmproj-GLM-5.3-Flash-Q8_0.gguf	652,829,376	0.61	Vision projector. Extra.
mtp-GLM-5.3-Flash-Q4_0.gguf	4,947,287,104	4.61	MTP sidecar. Pull #27917 still open.
mtp-GLM-5.3-Flash-Q8_0.gguf	9,284,197,440	8.65	MTP sidecar, larger.

There is no Q4_K_M, no 1-bit, and no 3-bit in this repo. The v0.6.0 server README says --hf-repo defaults to Q4_K_M, then falls back to the first file if that quant is missing. Pin :Q2_K or :Q4_K. Do not accept the fallback.

A secondary writeup called these files “around 150 GB.” Neither shard pair is 150 GiB. Q2_K is 124.62 GiB (133.81 decimal GB). Q4_K is 168.92 GiB (181.37 decimal GB).

Unsloth dynamic GGUF (separate publisher)

unsloth/GLM-5.3-Flash-GGUF, tree sums on 2026-10-06. Unsloth labels these in decimal GB. The byte sums match those labels. GiB is smaller, and GiB is the unit that compares to “128 GB” memory when the machine is counting gibibytes. The UD-IQ1_S header read the same day also says architecture glm5-next.

Quant	Bytes	GiB	Decimal GB (Unsloth’s label)
UD-IQ1_S	93,087,245,088	86.69	93.09
UD-IQ1_M	97,579,344,672	90.88	97.58
UD-IQ2_XXS	101,844,951,904	94.85	101.84
UD-Q2_K_XL	108,720,071,552	101.25	108.72
UD-IQ3_XXS	120,367,571,840	112.10	120.37
UD-Q3_K_XL	147,535,922,080	137.40	147.54
UD-IQ4_XS	156,822,111,200	146.05	156.82
UD-Q4_K_XL	199,707,321,472	185.99	199.71
Q8_0	340,981,966,208	317.56	340.98
BF16	641,641,064,288	597.57	641.64

Unsloth’s hardware table, still on the docs page fetched 2026-10-06, is total memory (RAM + VRAM, or unified memory), not file size: 1-bit 100 GB, 2-bit 115 GB, 3-bit 128–150 GB, 4-bit 162–210 GB, 8-bit 350 GB, BF16 650 GB. Their 3-bit demo file is UD-IQ3_XXS because they say it fits 128 GB devices (Mac or NVIDIA DGX Spark). That is a vendor estimate. LocalRig has not loaded the GGUF.

Context on the vendor cards is 1,048,576 tokens. KV cache sits on top of the weight file. A machine that just barely holds weights at short context does not hold a filled 1M window. Use the VRAM calculator only as a rough KV overlay.

What hardware fits which file?

Fit here means the weight file versus the memory label. It is not a measured load, and it is not tok/s. “Not yet measured” is the speed cell for every row.

Machine you have	ggml-org Q2_K (124.62 GiB)	ggml-org Q4_K (168.92 GiB)	Unsloth UD-IQ1_S (86.69 GiB)	Unsloth UD-IQ3_XXS (112.10 GiB)
24 GB card (3090 / 4090)	No	No	No	No
48 GB (two 3090s)	No	No	No	No
One 80 GB GPU	No. Weights exceed the card.	No	Weights exceed 80 GiB.	No
128 GB unified (Mac / DGX Spark)	Weights are 124.62 GiB. That is 5.41 GiB over 128 decimal GB (119.21 GiB) and leaves 3.38 GiB if the label is 128 GiB, before the OS, KV, and the 0.61 GiB mmproj. Not a measured fit.	No. Weights exceed both readings of 128 GB.	Unsloth’s table says 100 GB total memory. Stock v0.6.0 load of this file is not yet measured.	Unsloth says this is the 128 GB demo quant. Stock v0.6.0 load is not yet measured.
256 GB unified	Weights are under the label. KV at long context is not yet measured.	Same. Under the label, KV unmeasured.	Vendor 1-bit band.	Vendor 3-bit band, with memory left over relative to 128 GB.

Streaming a larger file from SSD is a different claim from this table. DwarfStar’s measured GLM 5.3 Flash Q4_K is 177.77 GiB on a 128 GB M5 Max, not a resident load of the ggml-org shards. See the Mac table below and SSD expert streaming on a Mac.

vLLM’s recipe is a different stack (safetensors, not these GGUFs). It lists vram_minimum_gb: 386 for default FP8 zai-org/GLM-5.3-Flash, 229 for the NVFP4 checkpoint (Blackwell, per the recipe), and 772 for BF16. Those are recipe figures dated in the file as updated 2026-09-18, not LocalRig measurements. None of them is a 24 GB or 128 GB home box.

NVIDIA DGX Spark on Amazon → is a 128 GB unified purchase. It matches Unsloth’s 3-bit claim, not a measured ggml-org Q2_K load, and not a 24 GB stand-in. Spark versus two 3090s: DGX Spark vs 2× RTX 3090. Browse used RTX 3090 24GB on eBay → only if you are shopping 27B-class models.

If your silicon is a single 24 GB card, the dense daily driver in this niche is Qwen 3.8 27B Q4.

Does Ollama, MLX, or vLLM support it, and in which version?

Three different answers. None of them is a LocalRig benchmark.

Ollama — cloud tag only, checked 2026-10-06. The library page documents ollama run glm-5.3-flash:cloud and says the model is hosted in the United States and Europe. The registry manifest for glm-5.3-flash:cloud returns 200. The manifest for glm-5.3-flash:latest returns manifest unknown. There is no local quant tag to pull. A cloud tag is an API call. It does not answer whether your GPU can hold the GGUF.

MLX — mlx-vlm, not a LocalRig file-size table. mlx-vlm v0.6.17, published 2026-08-26, lists “GLM-5.3-Flash support” (pull #2030). The file mlx_vlm/models/glm5_next/language.py is present on that tag and still present on tag v0.7.6 (published 2026-10-05). It is absent on v0.6.16. mlx-lm’s newest GitHub release as of the 2026-10-06 check is v0.31.3, published 2026-04-22, which predates this model. That mlx-lm tag was not re-checked on 2026-10-21. LocalRig still has no MLX quant-size table for this SKU.

MLX-Serve is a different project, and its speed line is a vendor claim. v26.10.3-pre-release.1, published 2026-10-20, says GLM-5.3-Flash reads experts from the SSD and runs on a 128 GB Mac, “writes about 35 tokens a second,” keeps “its full million-token context,” and starts a short question in about 2.5 seconds instead of 7. The sentence does not name the chip, the quant, or the prompt length. It is not mlx-lm, not mlx-vlm, and not a LocalRig run. The stable tag v26.10.2 (2026-10-12) lists GLM-5.3-Flash as running locally and does not contain the 35 tok/s sentence.

vLLM — release notes and the recipe disagree on how you install it. v0.30.0, published 2026-09-22, lists GLM-5.3-Flash (pull #53906) under new models and says the glm5next NVIDIA subtree is packaged in wheels (pull #55214). v0.31.0, published 2026-10-05, lists further GLM-5.3-Flash kernel work, so the model is still in that tag’s notes. The official recipe, date_updated: 2026-09-18, still sets min_vllm_version: "0.29.0", nightly_required: true, pip: false, and tells NVIDIA readers to use vllm/vllm-openai:glm53-flash until support lands in the standard image. That recipe date is older than the v0.30.0 notes. LocalRig has not installed a wheel or the special image and has not served the checkpoint. Treat 386 GB as the recipe’s FP8 floor, not as proof that pip install vllm==0.30.0 boots the model on a home GPU.

How fast is it? Which numbers are public, and on what hardware?

No number below was measured on llama.cpp v0.6.0. LocalRig has no first-party row.

Unsloth (Sep 4, 2026 note: “3.3× faster” plus MTP) reports UD-IQ1_S on 1× B200, MTP off first. Those notes still tell the reader to build Unsloth’s fork. They are vendor fork-era figures, not a stock v0.6.0 row:

Test	Baseline tok/s	Optimized tok/s
pp512	1121.80	1122.0
tg32	62.79	63.10
tg32 @ 4096	53.52	59.50
tg32 @ 16384	41.02	57.99
tg32 @ 65536	20.66	48.99

Accessed again 2026-10-06. They are not RTX 3090, RTX 4090, Spark, or Mac numbers. With MTP, Unsloth shows further gains (example: prompt 4096, MTP off 58.6 vs n=2 86.5 tok/s) and warns that more draft tokens can slow the run. Copy the hardware label with the number. MTP is not in the v0.6.0 tag. Pull #29928 merged 2026-10-07 and is listed in v0.7.0. Do not paste the Unsloth MTP column, or the llamaperf 61 tok/s row below, onto the v0.6.0 tag.

Two named Spark packets are older than the tag. They are community-cited, not LocalRig-measured, and they are not an A/B.

Source	Date	Hardware	Runtime	Quant	What was measured	tok/s
Weschera	2026-08-27	1× DGX Spark GB10, 121 GiB UMA	llama.cpp on eauchs pull #27752 (--spec-type draft-mtp). That pull closed unmerged on 2026-10-01.	Unsloth UD-Q2_K_XL (109 GB in that repo; Hugging Face sum is 108.72 decimal GB)	Decode at 32K context (temp 0, warmed). 262,144 boots with q8_0 KV; that is allocation, not this table.	Structured 33.8 · JSON 26.6 · code 25.0 · prose 17.9
gitcommit90	2026-09-03	1× DGX Spark, 128 GB UMA	vLLM TP1 + DFlash2 K5 (default). Community recipe, not the upstream FP8 recipe.	Turboderp EXL3 2.05 (85.23 GB)	Short C1 decode (400 out). Long-context column in that repo is prefill at 8k/16k/100k, not filled-262k decode. DFlash2 is CC BY-NC-ND 4.0.	Prose 29.9 · code 40.1 · structured 53.5

Do not average those rows. Do not paste them onto a Mac or onto v0.6.0.

Mac numbers, checked 2026-10-21

These are not LocalRig measurements, and they are not one test. The file-size tables above were not re-fetched on this pass. DwarfStar’s Q4_K file is 177.77 GiB, which is not the ggml-org Q4_K sum of 168.92 GiB. Mechanism and the longer comparison: SSD expert streaming on a Mac.

Source	Date	Label	Hardware	What was reported
DwarfStar SSD streaming	2026-09-06	Project measurement. Three-run median. SSD streaming, automatic cache, no speculative decoding.	128 GB M5 Max	GLM 5.3 Flash Q4_K, 177.77 GiB. 2K prompt, 1K append, 128 tokens. Prefill 121 then 104 t/s. Generation after each frontier 11.9 and 14.9 t/s. First-token wait included.
MLX-Serve v26.10.3-pre-release.1	Published 2026-10-20	Vendor pre-release claim. Chip, quant, and prompt not stated.	”a 128 GB Mac”	Experts read from SSD. About 35 tokens a second. Short-question start about 2.5 s instead of 7. Million-token context claimed in the same bullet.
llamaperf d9a97350	Oct 11, 2026	Community-cited single report. Quant not stated. Resident, not SSD streaming. llama.cpp version not stated. Reddit source URL returned HTTP 403 on 2026-10-21.	M5 Ultra Mac Studio, 256 GB	39.7 tok/s without MTP. 61.0 tok/s with the built-in MTP head, 3 draft tokens. 262,000 context, about 15 GiB spare.
LLMCheck	Updated 2026-10-17	Index estimate, bandwidth model, factor 0.42. Not a measurement.	M5 Max 128 GB at their 3-bit cost; M5 Ultra at their 4-bit cost	About 27 tok/s estimated on the M5 Max. About 35 tok/s estimated on the M5 Ultra. That 35 is not the MLX-Serve sentence.

LLMCheck also summarizes a community oMLX run at about 13 tok/s (2.5-bit, about a third of experts streamed, 1K context, 128 GB M5 Max). The linked oMLX page returned HTTP 403 on 2026-10-21, so that 13 is not repeated here as a confirmed row. First-party Mac tok/s is still a shortfall.

Should I rent an A100 or H100 to try it?

A single 80 GB A100 or H100 does not hold any weight file in the tables above. ggml-org Q2_K is 124.62 GiB. Unsloth UD-IQ1_S is 86.69 GiB. Unsloth’s 3-bit band is 128–150 GB. The vLLM recipe’s FP8 floor is 386 GB, which is a multi-GPU number.

Vultr’s catalog is datacenter GPUs (A100, H100, L40S), not consumer 4090s. Check current multi-GPU inventory on Vultr GPU Cloud if you are actually shopping that class. Do not treat the link as “this card runs GLM-5.3-Flash.” For a 24 GB experiment on a different model, use RunPod or Vast.ai. Full IaaS positioning: Vultr GPU cloud review.

How do I load it on stock llama.cpp v0.6.0?

Use the v0.6.0 tag, published 2026-10-05. The v0.6.0 README’s installed command is llama serve. The same tag’s server README documents llama-server -hf user/model:quant and says the optional quant defaults to Q4_K_M. This repo has Q2_K and Q4_K only. Pin the quant. The model card’s one-liner omits it:

# Documented on the ggml-org model card. Does not name a quant.
llama serve -hf ggml-org/GLM-5.3-Flash-GGUF

# v0.6.0 server README shape, quant pinned to a file the repo actually has.
llama serve -hf ggml-org/GLM-5.3-Flash-GGUF:Q2_K

Q4_K is the same command with :Q4_K. The GGUF header stores sampling temperature 1.0 and top_p 0.95 (float32). The server README says -hf also downloads mmproj when one is present (0.61 GiB here) unless you pass --no-mmproj. LocalRig has not run these commands. If the installed binary from a source build is still named llama-server, the server README’s flag is the same -hf form.

Pre-built binaries and the llama.app install script are both listed in the v0.6.0 README. Either one has to be v0.6.0 or newer. An older tag will not have this architecture.

Related runtime overview, not a version pin for this model: Ollama vs llama.cpp vs vLLM. The 24 GB GLM Flash in this batch is GLM-4.7-Flash (30B-A3B, Unsloth ~18 GB 4-bit).

Historical note: Unsloth’s fork (do not use for a new install)

On 2026-09-10 this page told readers to build Unsloth’s branch glm5next/upstream, because that is what their guide printed and upstream had no tag for the architecture yet. Fetched again 2026-10-06, Unsloth’s guide still contains:

git clone --branch glm5next/upstream https://github.com/unslothai/llama.cpp

and still points the 3-bit demo at UD-IQ3_XXS. Their Hugging Face card still says to use llama.cpp pull #27754, which closed unmerged on 2026-10-01. That snippet is stale vendor copy next to the v0.6.0 tag. The dynamic files’ architecture string is now glm5-next, the same string as the ggml-org files. Whether stock v0.6.0 loads Unsloth’s dynamic quants is not yet measured. If a dynamic file fails on v0.6.0, the failure belongs in a later update. It is not a reason to build a closed pull or the old fork branch.

Unsloth Desktop remains their own app path. This page does not re-test it.

Who this is NOT for

• Anyone who wants LocalRig tok/s on llama.cpp v0.6.0. There is no first-party row. The B200 table is Unsloth’s fork-era bench. The Spark rows are community recipes on a closed llama.cpp pull and on a community vLLM EXL3 build. The Mac rows added 2026-10-21 are DwarfStar, MLX-Serve, and llamaperf, and they are not a v0.6.0 measurement.
• Buyers with a 12–24 GB Ampere or Ada card. Wrong SKU. Use Qwen 3.8 27B.
• People collapsing this into GLM-5.2. Different model, different memory class. Keep the 5.2 page for the flagship.
• Anyone treating “around 150 GB” as the official file. The tree sums are 124.62 GiB and 168.92 GiB.
• Anyone assuming one 80 GB A100, or a 128 GB Mac, holds ggml-org Q2_K resident. 124.62 GiB of weights is not an 80 GB card and is not a measured resident 128 GB load. DwarfStar’s 11.9 / 14.9 tok/s row is a different file, 177.77 GiB, streamed from SSD.
• Anyone who still clones glm5next/upstream or pull #27752 / #27754. Those were the pre-tag paths. The tag is v0.6.0. The merged pull is #27773.

Methodology

• Runtime pin: llama.cpp release notes for tag v0.6.0, published 2026-10-05, plus pull #27773 merged 2026-09-30. Closed unmerged pulls #27752 and #27754 checked 2026-10-06. MTP pull #27917 still open the same day.
• Official file sizes: Hugging Face tree API for ggml-org/GLM-5.3-Flash-GGUF, accessed 2026-10-06. Shard bytes summed, then divided by 1024³ for GiB. Architecture, context length, and sampling floats read from the Q2_K first-shard GGUF header the same day. Unsloth UD-IQ1_S header checked for general.architecture only.
• Unsloth sizes and memory bands: docs page and Hugging Face tree sums, both accessed 2026-10-06. Decimal GB labels match the byte sums divided by 1000³. Memory bands are vendor estimates.
• Ollama: library page and registry manifests for :cloud (HTTP 200) and :latest (manifest unknown), 2026-10-06.
• MLX: mlx-vlm release notes v0.6.17 (published 2026-08-26) and presence of glm5_next on tags v0.6.17 and v0.7.6 versus absence on v0.6.16. mlx-lm newest release tag observed: v0.31.3, published 2026-04-22.
• vLLM: release notes v0.30.0 (published 2026-09-22) and v0.31.0 (published 2026-10-05), plus recipes YAML date_updated: 2026-09-18 fetched 2026-10-06. The recipe’s nightly / pip flags are quoted because they conflict with the later release notes. Not installed here.
• Speed numbers: Unsloth llama-bench on 1× B200, UD-IQ1_S, fork-era notes still on the docs page as of the 2026-10-06 fetch. Spark: Weschera on closed pull #27752 (2026-08-27); gitcommit90 vLLM EXL3 2.05 (2026-09-03). Not LocalRig-measured. Not transferable to v0.6.0, a 3090, or a 4090.
• Mac speed, fetched 2026-10-21: DwarfStar SSD streaming, GLM 5.3 Flash Q4_K 177.77 GiB, 128 GB M5 Max, generation 11.9 and 14.9 t/s, dated 2026-09-06. MLX-Serve v26.10.3-pre-release.1, published 2026-10-20, about 35 tok/s on a 128 GB Mac, vendor claim, chip and quant unstated. llamaperf d9a97350, Oct 11 2026, 39.7 tok/s plain and 61.0 with MTP on a 256 GB M5 Ultra, quant unstated, resident. LLMCheck estimates are labeled estimates. Not first-party. Not one test.
• MTP status, fetched 2026-10-21: pull #27917 still open. Pull #29928 merged 2026-10-07 and is listed in llama.cpp v0.7.0, published 2026-10-15. Not run here.
• LocalRig first-party: none for GLM-5.3-Flash as of 2026-10-21. No engine.db benchmark row. File-size sums in the tables above remain the 2026-10-06 fetch.

Sources

• llama.cpp v0.6.0 — published 2026-10-05. GLM-5.3-Flash / GLM5-Next listed under new models, pull #27773. README install command llama serve. Accessed 2026-10-06.
• llama.cpp v0.6.0 server README — --hf-repo quant default Q4_K_M, mmproj auto-download. Accessed 2026-10-06.
• Pull #27773 — merged 2026-09-30. The support that shipped.
• ggml-org/GLM-5.3-Flash-GGUF — created 2026-10-03. Q2_K 124.62 GiB, Q4_K 168.92 GiB, mmproj and MTP sidecars. Tree API accessed 2026-10-06.
• Unsloth GLM-5.3-Flash guide — memory table, dynamic file labels, B200 llama-bench, and the fork clone that is now a historical note. Accessed 2026-10-06.
• unsloth/GLM-5.3-Flash-GGUF — dynamic quant byte sums, accessed 2026-10-06. Card still links closed pull #27754.
• Ollama glm-5.3-flash — :cloud only, accessed 2026-10-06.
• mlx-vlm v0.6.17 — published 2026-08-26. GLM-5.3-Flash support. Module still on v0.7.6 (2026-10-05).
• vLLM v0.30.0 — published 2026-09-22. Lists GLM-5.3-Flash. v0.31.0 published 2026-10-05.
• vLLM recipe — date_updated: 2026-09-18. nightly_required: true, pip: false, FP8 386 GB. Accessed 2026-10-06.
• Weschera/glm53-flash-one-spark — 1× Spark, UD-Q2_K_XL, 32K decode, 2026-08-27, on closed pull #27752.
• gitcommit90/glm-5.3-one-spark — 1× Spark vLLM EXL3 2.05 + DFlash2, 2026-09-03.
• DwarfStar SSD streaming — 128 GB M5 Max, GLM 5.3 Flash Q4_K 177.77 GiB, generation 11.9 and 14.9 t/s, 2026-09-06. Fetched 2026-10-21.
• MLX-Serve v26.10.3-pre-release.1 — published 2026-10-20. About 35 tok/s is a vendor claim.
• llamaperf report d9a97350 — Oct 11, 2026. 256 GB M5 Ultra, 39.7 / 61.0 tok/s, quant not stated.
• LLMCheck — updated 2026-10-17. Index estimates, not measurements.
• llama.cpp v0.7.0 — published 2026-10-15. MTP via #29928, merged 2026-10-07. #27917 still open the same check.
• SSD expert streaming on a Mac — mechanism, trade-offs, and the comparison table.
• Can I run GLM-5.2 locally? — different SKU.
• Can I run GLM-4.7-Flash locally? — 24 GB GLM Flash, not this SKU.
• Can I run Qwen 3.8 locally? — 24 GB dense path.
• DGX Spark vs 2× RTX 3090 — 128 GB unified vs 48 GB split.
• Vultr GPU Cloud — datacenter inventory, not a promised single-GPU fit.