Four-bit weights leave 6 GB on a 24 GB card for KV cache and vision tensors

Start at the load path. Quantize to 4 bits with AWQ, EXL2 or GGUF Q4_K_M and the weights sit at 16 to 18 GB, leaving 6 to 8 GB free on a 24 GB card ^[16]. Everything the model does at runtime comes out of that remainder. The dev.to comparison tells teams running Qwen on video to reserve at least 4 GB of buffer for the spike during visual feature extraction ^[18]. Subtract that and the KV cache is working with 2 to 4 GB ^[6].

That figure decides your serving profile. It is also why the context spec does not travel. Qwen3.8-27B advertises 262,144 tokens natively ^[9]. The same article says sequences past 32,768 tokens on a single 24 GB card need FlashAttention-2 plus paged KV cache quantization in FP8 or INT4 ^[17]. Eight to one ^[5]. The default ceiling on your box is an eighth of the model card's headline. Closing the gap means editing launch flags, not buying a licence.

Quantization is doing the heavy lifting here; parameter count is not. Unquantized, both models want 54 to 60 GB and therefore more than one GPU ^[15]. Four-bit weights are roughly a 3.3x cut on that requirement ^[4].

The two labs answer the serving question differently. Qwen3.8-27B arrived supported on vLLM, SGLang and llama.cpp behind standard OpenAI-compatible endpoints ^[19]. Muse Glimmer instead ships explicit 24, 32 and 64 GB deployment packages ^[13], which is a profile you adopt rather than a runtime you already operate. The copy of the article available to us breaks off mid-item where Muse Glimmer's runtime support would be listed ^[2], so treat parity there as unverified.

No benchmark scores appear for either model in that text ^[3]. If a table shows up later, it transfers to your hardware only when the quantization format, the KV cache dtype and the image resolution match what you actually serve, and those three choices are exactly what the 2 to 4 GB above constrains.

The licence delta is real and boring, which is the best kind. Qwen3.8-Max, announced 3 August, needs an explicit agreement once a model-as-a-service or AI assistant business passes US$50m in annual revenue ^[10], and Moonshot AI's Kimi K3 gates above US$20m ^[11]. A vendor sitting between those two lines is clear of one and captured by the other ^[7]. Neither of the ~30B releases carries a threshold at all ^[14]. For an on-premise appliance that deletes a recurring audit. It is not a latency problem.

One caution on the evidence. The comparison credits its architecture and release details to unnamed independent audits and unnamed release tracking ^[1]. Read the model card yourself before shipping an appliance on the strength of it.

My read, with my context stated: one 4090, agent loops, short prompts and many turns. Muse Glimmer's 24 GB package and the KV budget matter more to me than a 262K window I cannot fill ^[13][6]. If the workload were hour-scale video understanding, only one of the two claims it, and it comes with a 4 GB reservation attached ^[9][18].