Last updated: October 2026
Key Takeaways
- Yes, on 128GB, at 1-bit to 3-bit. Measured runs reach about 15 to 18 tokens per second at typical context lengths on a Strix Halo mini PC and a DGX Spark. The real 1-bit file is 93GB, larger than our launch estimate, which rules out 128GB Strix Halo machines running Windows.
- Mainline llama.cpp has supported GLM-5.3-Flash since build b11279 on September 30, 2026. Ollama's library offers only a cloud tag and LM Studio launched it as a cloud model, so a fully local run means llama.cpp, Unsloth's tools, or MLX.
- It is the highest-scoring open-weights model a 128GB machine can load, and it is MIT-licensed. Full 4-bit quality needs about 200GB of memory.
Yes. GLM-5.3-Flash runs on a 128GB machine at 1-bit to 3-bit quantization, now in mainline llama.cpp, and the measured setups below reach about 15 to 18 tokens per second. The 4-bit file, which keeps about 92 percent of full-precision accuracy, needs roughly 200GB. Machines with 64GB or less have no path.
This page launched on August 26 with estimated file sizes. The sizes below are the real files, the speeds come from named testers, and every status was checked on October 5, 2026.
What Changed Since Launch
Four things moved after August 26: the files landed, mainline llama.cpp merged support, the flagship's weights shipped under a different license, and people measured speeds.
- Late August: flagship GLM-5.3's open weights went up on Hugging Face under a custom GLM-5.3 license, not MIT. Our GLM-5.3 open-weights ledger covers its clauses.
- September 4: Unsloth added multi-token prediction (MTP) to its llama.cpp branch, a speculative-decoding speedup it reports at up to 3.3x at long context.
- September 29: a public Strix Halo measurement list recorded 17.6 tokens per second at 1-bit.
- September 30: llama.cpp merged GLM-5.3-Flash support in pull request #27773, first shipped in build b11279, covering text and image input. MTP for this model is still a separate, unmerged pull request.
- October 1 and 5: two follow-up fixes to the k-pool attention cache this architecture uses landed in mainline. Use a build from October 5 or later.
What We Got Wrong
Our launch estimate put the 1-bit file at 75-80GB. The real UD-IQ1_S file is 93.09GB. That error changes one verdict: a 128GB Strix Halo machine on Windows, which caps the GPU at 96GB, cannot hold the file with working context. Our flagship estimates were low too: its smallest GGUF is 217GB, not 176GB.
What GLM-5.3-Flash Is
GLM-5.3-Flash is Z.ai's 320-billion-parameter mixture-of-experts model with 18 billion parameters active per token, image and video input, a 1M-token context window, and an MIT license, per its model card on Hugging Face. Per Z.ai's documentation, its hybrid sparse-and-linear attention cuts attention compute 3.01x and KV cache 4.44x against flagship GLM-5.3. Before launch it ran anonymously as "ox-alpha" on OpenRouter and OpenCode.
On the Artificial Analysis Intelligence Index (v4.3.2, checked October 5), Flash scores 42, fourth among open-weights models. The three above it are Xiaomi's MiMo-V2.6-Pro at 46 (1.02T parameters), flagship GLM-5.3 at 45 (753B), and Kimi K3 at 44 (2.8T; see our Kimi K3 reality check). None of those three fits on a 128GB machine at any quant. Its closest same-size rival, Xiaomi's MiMo-V2.6-Flash (309B total, 15B active, MIT), scores 38. The index tests the full-precision model through an API; the quants a 128GB machine runs give up some of that quality.
GGUF Sizes: Every Quant, Measured
The smallest Unsloth file is 93GB and its best 4-bit file is 200GB; every quant is listed below with the smallest memory class that holds it.
| Quant | File size | Top-1 accuracy | Smallest memory that holds it |
|---|---|---|---|
| UD-IQ1_S (1-bit) | 93.09GB | 70.9% | 128GB on Linux, or a 128GB Mac with the GPU limit raised |
| UD-IQ1_M (1-bit) | 97.6GB | Not published | 128GB on Linux |
| UD-IQ2_XXS (2-bit) | 102GB | Not published | 128GB on Linux, short context |
| UD-Q2_K_XL (2-bit) | 108.72GB | 78.3% | 128GB on Linux, tight |
| UD-IQ3_XXS (3-bit) | 120.37GB | 81.6% | Loads on 128GB with almost no context; 192GB is comfortable |
| UD-Q3_K_XL (3-bit) | 148GB | Not published | 192GB |
| UD-IQ4_XS (4-bit) | 157GB | Not published | 192GB |
| UD-Q4_K_XL (4-bit) | 199.71GB | 92.2% | 256GB |
| UD-Q5_K_XL (5-bit) | 240.31GB | 94.4% | More than 256GB |
| UD-Q6_K_XL (6-bit) | 291.83GB | 95.2% | More than 256GB |
| Q8_0 (8-bit) | 341GB | Not published | 384GB or more |
Sizes and accuracy from Unsloth's GGUF repository and Unsloth's GLM-5.3-Flash guide, checked October 5, 2026. Top-1 accuracy is Unsloth's own measure relative to the full-precision model; "Not published" means its table omits that quant. Add about 1.1GB for the vision projector (mmproj) if you want image input.
Two community files are built specifically for 128GB machines: aj9o9's GGUF set offers a 3-bit AJ-IQ3_XXS at 112.40GB and a 2-bit AJ-IQ2_XXS at 87.35GB. On a 128GB machine the 3-bit one leaves about twice the free memory of Unsloth's 120GB 3-bit file, enough for real context.
Which Quant for Your Memory
Pick the largest file that still leaves 10GB to 15GB free for context and the operating system.
- 128GB on Linux (Strix Halo with the memory unlock, DGX Spark, ASUS Ascent GX10): AJ-IQ3_XXS for quality with room to work, UD-Q2_K_XL as the Unsloth choice, or UD-IQ1_S for long context.
- 128GB on Windows: none. The 96GB GPU cap is below every file plus context.
- 128GB Mac: UD-IQ1_S once you raise macOS's GPU wired-memory limit; 2-bit is borderline.
- 192GB: UD-IQ3_XXS or UD-IQ4_XS with comfortable context.
- 256GB: UD-Q4_K_XL, the quality tier; on a Mac, the 204GB MLX 4-bit build.
What Low-Bit Costs
Unsloth's table is the only per-quant quality data published for Flash, and it is blunt: the 1-bit file keeps about 71 percent of full-precision top-1 accuracy, 2-bit about 78 percent, 3-bit about 82 percent, and 4-bit 92 percent. The biggest single step is from 3-bit to 4-bit, and that step is also the line between a 128GB machine and a 256GB one. Treat a 128GB run as a capable preview of the model, not the model the benchmark charts describe.
The 128GB Line, Measured
Three 128GB setups have published GLM-5.3-Flash speeds, and without MTP they land between about 10 and 18 tokens per second depending on quant and context length.
| Hardware | Quant and context | Measured speed | Source |
|---|---|---|---|
| Ryzen AI Max+ 395 (128GB), Linux, Vulkan | UD-IQ1_S, short prompt | 17.6 tok/s output, 145 tok/s prompt; 15.5 tok/s at 4K depth; about 25 tok/s with MTP | strix-halo-llm-speeds |
| Ryzen AI Max+ 395 (128GB), Linux, Vulkan | AJ-IQ3_XXS, 250K context | 15.6 tok/s output, 109 tok/s prompt | ds4 issue #900 |
| Ryzen AI Max+ 395 (128GB), Linux, Vulkan | UD-Q2_K_XL, 500K context | 9.7 tok/s output | ds4 issue #900 |
| NVIDIA DGX Spark (128GB) | UD-Q2_K_XL, 65K context | 17.7 tok/s single stream; 64 tok/s across 8 streams; 110 of 121GiB used | Classmethod's DGX Spark test |
| 128GB Mac | UD-IQ1_S | No published measurement | Not applicable |
Each figure is its tester's own measurement, taken on Unsloth's branch before the September 30 mainline merge; mainline speeds may differ. Checked October 5, 2026.
On Strix Halo, every measured run depends on Linux and one memory setting. Windows caps GPU-addressable memory at 96GB. Linux can hand the GPU about 120GB through the amdgpu.gttsize and ttm.pages_limit kernel parameters, leaving about 8GB for the operating system, per this Strix Halo setup guide. Our Ryzen AI Max+ 395 reality check covers the rest of the platform. The GMKtec EVO-X2 is the 128GB box this page's numbers were written around; the same chassis also ships with 64GB, which cannot hold any GLM-5.3-Flash file, so confirm 128GB in the listing title.
Check Price on Amazon: GMKtec EVO-X2 (128GB)
NVIDIA's DGX Spark is the other measured 128GB box. Classmethod's DGX Spark test ran the 2-bit file at 65K context with about 11GiB to spare, and its code-generation tests passed. Buy the NVIDIA-branded unit, which is 128GB; a 64GB configuration exists only through partners.
Check Price on Amazon: NVIDIA DGX Spark (128GB)
On a Mac, the 128GB class holds the 1-bit file once the GPU wired-memory limit is raised, but nobody has published a measured speed. For 4-bit quality on Apple hardware, the MLX community 4-bit build is 204GB and needs a 256GB Mac. Apple's Mac Studio page lists the M5 Ultra at up to 512GB, with the 512GB option arriving in late October; our M5 Ultra wait-or-buy guide covers which tier earns its price.
Can an RTX 3090 run it? Not alone: 24GB holds the attention layers and cache, not the 93GB-plus of weights. Paired with 192GB or more of system RAM it runs the 3-bit and 4-bit files, and Unsloth's guide sizes its quants for exactly this RAM-plus-VRAM split. No measured Flash speed for that build has been published yet. Our best local AI models by VRAM guide covers the card choice.
Check Price on Amazon: RTX 3090 Founders Edition 24GB (Renewed)
Runtime Ledger, October 5
Mainline llama.cpp is now the dependable local route; the two one-click apps offer cloud versions, not local ones.
| Runtime | Status, October 5, 2026 | What it means |
|---|---|---|
| llama.cpp mainline | Supported since b11279 (September 30); text and image input; MTP not yet merged | The default local route. Use an October 5 or later build for the k-pool fixes |
| Unsloth llama.cpp branch | Supported, with MTP since September 4 | Fastest measured route: about 25 tok/s at 1-bit on Strix Halo |
| Ollama | Library lists only glm-5.3-flash:cloud | Runs on Ollama's servers in the US and Europe. Not local |
| LM Studio | Launched as a cloud model in LM Studio Bionic; no local support announced | Cloud by default. Check that its bundled llama.cpp runtime is b11279 or newer before trying a GGUF |
| MLX (Apple) | Supported in mlx-vlm, not in mlx-lm | The 204GB 4-bit build needs a 256GB Mac |
| vLLM and SGLang | Supported per Z.ai's model card | Server-class memory, not a desktop route |
Checked October 5, 2026 against the llama.cpp repository history, Ollama's library page, LM Studio's catalog, the MLX model card, and the Z.ai model card.
Settings That Matter Locally
- Reasoning effort: Z.ai's documentation recommends reasoning_effort set to max. Locally that is expensive: Classmethod's DGX Spark test took 29 minutes on ten tasks at max and 98 seconds at low, with comparable quality. Use low or high for interactive work.
- Sampling: temperature 1.0 and top_p 0.95, per Z.ai and Unsloth.
- MTP: on Unsloth's branch, keep the draft length at 2; Unsloth reports that longer drafts slow it down.
Storage and Download Math
Each quant is a 93GB-to-200GB download, so comparing three files costs more than 300GB of disk. At 100 Mbps, a 109GB 2-bit file takes about two and a half hours; gigabit fiber takes about 15 minutes. A 4TB NVMe drive holds a quant ladder plus your existing library.
Check Price on Amazon: Samsung 990 EVO Plus 4TB NVMe
Pricing a 192GB-plus build? Our RAM shortage guide explains why buying machines with memory installed beats waiting.
If You Are Under the Floor
Below about 96GB of usable memory there is no GLM-5.3-Flash path. DeepSeek's newer model is not the 128GB answer either: our V4.1-Flash reality check found it needs two DGX Sparks. DeepSeek V4-Flash 0731 still loads on one 128GB machine in mainline llama.cpp and LM Studio; our V4-Flash reality check has the tier map. Below 128GB, what AI models actually run at home sorts the field by memory tier, and our mini PC guide for local AI maps always-on boxes.
The Open-Infrastructure Read
The license is half the case for owning this model. Flash ships under plain MIT: no user-count clause, no revenue threshold, no regional carve-out. Flagship GLM-5.3's custom license adds conditions.
The second half is where your prompts go. Both one-click apps shipped GLM-5.3-Flash as a hosted service first: Ollama's only tag runs on Ollama's servers, and LM Studio's launch support ran in its cloud. Their zero-data-retention promises are policies, not architecture. At $0.15 per million input tokens and $0.50 per million output tokens on Artificial Analysis's October 5 listing, running Flash at home is not a cost play. It is about who holds the weights and who sees the prompts. A model on your own disk, behind your own router, is infrastructure you control; our GLM-5.1 analysis makes the same case about the lab's own compute.
Frequently Asked Questions
Is GLM-5.3-Flash free for commercial use?
Yes. GLM-5.3-Flash is released under the MIT license, which allows commercial use, modification, and redistribution with attribution, with no revenue threshold or user-count clause. Flagship GLM-5.3 is different: it uses a custom GLM-5.3 license with extra conditions. Confirm the LICENSE file in the official repository before building a product on either.
How big is the GLM-5.3-Flash GGUF?
Unsloth's files run from 93.09GB for the 1-bit UD-IQ1_S to 199.71GB for the 4-bit UD-Q4_K_XL, 341GB for Q8_0, and 641.64GB for full BF16. The 2-bit UD-Q2_K_XL is 108.72GB. Community builds go lower: a 2-bit AJ-IQ2_XXS is 87.35GB. Add about 1.1GB for the vision projector if you want image input.
Will GLM-5.3-Flash run on 64GB of RAM?
No. The smallest file we found, a community 2-bit build, is 87.35GB, which exceeds 64GB before any context is allocated. No configuration of GLM-5.3-Flash fits a 64GB machine. On that hardware, a smaller model that fits fully in memory is the better choice.
How fast is GLM-5.3-Flash on Strix Halo?
On a 128GB Ryzen AI Max+ 395 running Linux with the Vulkan backend, the 1-bit file measures about 17.6 tokens per second, rising to about 25 with multi-token prediction on Unsloth's branch. A community 3-bit file measures 15.6 tokens per second at 250K context. These runs need Linux, because Windows caps GPU memory at 96GB.
Does Ollama run GLM-5.3-Flash locally?
No. As of October 5, 2026, Ollama's library lists only glm-5.3-flash:cloud, which runs on Ollama's servers rather than your machine. For a local run, use mainline llama.cpp from build b11279 or later, Unsloth's tools, or MLX on a Mac. LM Studio also launched it as a cloud model, with no local support announced yet.
Can I run GLM-5.3-Flash on a Mac?
Yes, with limits. A 128GB Mac can hold the 93GB 1-bit GGUF once the GPU wired-memory limit is raised, but no measured speed has been published. For 4-bit quality, the MLX community build is 204GB and needs a 256GB Mac, running through mlx-vlm rather than mlx-lm.
What is the difference between GLM-5.3 and GLM-5.3-Flash?
They are separate models. Flagship GLM-5.3 has 753B parameters with 40B active, is text-only, and uses a custom license; its smallest GGUF is 217GB. GLM-5.3-Flash has 320B parameters with 18B active, accepts images and video, uses the MIT license, and has a 93GB 1-bit file. Flash scores 42 to the flagship's 45 on Artificial Analysis's index.
