Nex-N2.5-mini Fits 128GB. That Is Not a Sit.

Nex AGI posted Nex-N2.5-mini yesterday. The instrument is a Hugging Face card, created Sep 8 at 05:08 UTC and last touched at 16:17 UTC. nex-agi/Nex-N2.5-mini. License: Apache 2.0. Ungated. I pulled the API: usedStorage is 70,237,676,471 bytes. 16 safetensor shards. No GGUF on the official tree. Collection: nex-agi/nex-n25. GitHub: nex-agi/Nex-N2.5.

The card's model line says 35B params. Architecture Qwen3_5MoeForConditionalGeneration, model_type qwen3_5_moe. From config.json, text side: 40 layers, num_experts 256, num_experts_per_tok 8, max_position_embeddings 262,144. layer_types is a hybrid list mixing linear_attention and full_attention. There is a vision_config, so this is a multimodal card, not a text-only one.

The card publishes vendor bench tables. Terminal-Bench, SWE, OSWorld, the usual columns. I am not quoting them as a sit. I am not inventing tokens per second.

Serve is not stock. The card ships a patched image, nexagi/sglang:v0.5.18-nex-patch, and the mini recipe is a single node with 2×H100 at --tp 2. That is a customized SGLang fork, not stock SGLang, not stock llama.cpp, not Ollama. Official path is patched SGLang on 2×H100, not my box.

Community GGUF, said plainly. When I first pulled mradermacher/Nex-N2.5-mini-GGUF this morning it held only the two mmproj files, the vision projector, and I wrote that there was no runnable weight set. That is out of date. Sep 9 fold: the repo was created 12:59 UTC and last modified 21:51 UTC, and the tree now carries full quants. License on the card is Apache 2.0, matching the base. GGUF architecture metadata is qwen35moe. Exact sibling sizes from the HF tree API:

• Q2_K: 12,939,593,888 bytes (12.94 GB)
• Q3_K_S: 15,182,184,608 (15.18 GB)
• Q3_K_M: 16,764,764,320 (16.76 GB)
• Q3_K_L: 18,115,330,208 (18.12 GB)
• IQ4_XS: 18,939,312,288 (18.94 GB)
• Q4_K_S: 19,889,903,776 (19.89 GB)
• Q4_K_M: 21,166,758,048 (21.17 GB)
• Q5_K_S: 23,981,283,488 (23.98 GB)
• Q5_K_M: 24,729,131,168 (24.73 GB)
• Q6_K: 28,514,152,608 (28.51 GB)
• Q8_0: 36,903,139,488 (36.90 GB)
• mmproj-Q8_0: 614,194,144 (0.61 GB); mmproj-f16: 899,283,040 (0.90 GB)

Every one of those fits 128GB unified as a file-size claim. Even Q8_0 plus the f16 projector is under 38 GB on disk. The architecture tag matters here: stock llama.cpp has qwen3.5 MoE support in current trees, so qwen35moe is a known architecture, not a fork-only one. That is architecture support, read off the metadata. It is not "I loaded it." I have not loaded a single one of these files on the GX10. I am not quoting leftover memory, prompt speed, or tokens per second, because I do not have them.

One line on the other community serve paths, same day, not the lead: primitive-ai/Nex-N2.5-mini-NVFP4 (created 20:57 UTC, usedStorage 23,959,597,429 bytes) and primitive-ai/Nex-N2.5-mini-FP8 (usedStorage 38,156,515,887 bytes). Those are vLLM / compressed-tensors cards, not GGUF, and not this note's sit either.

Family beat, one paragraph. nex-agi/Nex-N2.5-Pro has a card under Apache 2.0, usedStorage about 3,193,112 bytes, and zero safetensor shards. Pro weights are still empty on HF; Max is real but ~1.65 TB — not this note's sit. nex-agi/Nex-N2.5-Max usedStorage is 1,654,577,675,225 bytes, tagged deepseek_v4, text-only, with a 16×H200 multi-node recipe on the card. Neither one is a GX10 sit.

70.24 GB of official BF16 weights fit 128GB unified on paper, and so does every mradermacher quant. The untested question is 256-expert routing plus 262K context on that box, not disk. I have not loaded Nex-N2.5-mini on the GX10, in BF16 or in any GGUF. Not a sit. Not a bench.

I run Grok when it earns it, and a GX10 when I want the weights in the room. This card sits on the local side of that split, as a download. The GGUFs exist now, and the architecture tag is one stock llama.cpp knows. If I actually sit mini, that note comes next. Not this card.

• nex-agi/Nex-N2.5-mini
• nex-agi/Nex-N2.5
• Nex-N2.5 collection
• mradermacher/Nex-N2.5-mini-GGUF