SummaryGLM-5.3-Flash is Z.ai's natively multimodal mixture-of-experts model with about 320B total and 18B active parameters, under the MIT license. Refusal-reduced variants include huihui-ai's abliterated GGUF, which edits layers 15 to 35 and no experts, and dealignai's FP8 release with an undisclosed weight edit. At this size, local use needs a multi-GPU machine.
What the base model is
GLM-5.3-Flash is, per its card, “the first natively multimodal model in the GLM-5 series”, with 320 billion total and 18 billion active parameters. Its configuration lists 288 routed experts with 8 used per token, and a hybrid architecture that combines linear attention with sparse attention to cut the cost of long contexts. Z.ai publishes it in FP8 under the MIT license, with a separate BF16 repository.
The card does not state a context limit, although its configuration allows 1,048,576 tokens and some of its evaluations were run “under 1M context”. Thinking depth is set with reasoning_effort, which accepts low, high and max; the card says it defaults to max when the parameter is missing or has any other value. That default matters for cost, because reasoning tokens are billed as output on most APIs.
The variants at a glance
huihui-ai’s abliterated GGUF says that only layers 15 to 35 were ablated, counting from zero, and that the other layers and all expert modules were left unchanged. The files are derived from Unsloth’s GGUF conversion, and the card notes that current llama.cpp releases support the model. As with huihui-ai’s other releases, the creator calls the method “a crude, proof-of-concept implementation”.
dealignai’s FP8 release describes a weight-level edit it calls “CRACK” without publishing the procedure, reports “0% refusals on HarmBench-320” in some reasoning modes, and says the edit was kept conservative so some refusals remain by design. Its card also carries the most detailed serving notes we have seen for this model: reasoning_effort values other than low and high silently become max, and at max a small max_tokens can return an empty reply as the model spends its budget thinking.
Which one for which need
- Self-hosting on a multi-GPU server: dealignai’s FP8 weights run in vLLM or SGLang at roughly the size of the original FP8 checkpoint, about 320 GB. huihui-ai’s GGUF files offer smaller quantizations for llama.cpp.
- Hosted: no provider in our data names either variant’s repo. MolochAI lists
glm-5.3-flash, described as a refusal-reduced build, without naming the weights. The standard model is available from many providers at low rates, but it is not refusal-reduced.
Limits and open questions
At about 320B parameters, this model does not fit on any single consumer GPU at any useful precision; the memory table below gives the arithmetic. Because of the hybrid sparse and linear attention, the VRAM estimator does not estimate its KV cache.
Neither variant publishes a quality measurement. Before relying on the dealignai build, read its serving notes; before relying on either for image input, test it, since vision quality after a language-layer edit is not documented.
Ordered as catalogued. Method, layers and claims are as each creator states them; Atlas has not tested any of these variants.
Estimate for GLM-5.3-Flash with 32,768 tokens of context, one sequence, FP16 KV cache and 10% overhead. Variants of the same architecture need about the same.