Spark X2.6 4B

4.1B parameter open-weight model. A compact general-purpose and agent model from XHToken's Spark line: 4.11B parameters (Hugging Face safetensors total), 36 layers, hidden size 2,560, 16 attention heads with 4 KV heads, a 131,072-token vocabulary and a claimed 1,048,576-token context (config.json and GGUF header read 2026-11-03). The GGUF header reports 4.45B because it stores the tied embedding matrix twice (131,072 x 2,560 = 335,544,320 parameters; the config sets tie_word_embeddings), so 4.1B is the model's size. Post-trained from Spark-X2.6-4B-Base on the same architecture as Spark X2.5, aimed at coding, agent workflows, long context and multilingual (English and Chinese) use, with selectable thinking levels. Apache 2.0. The vendor's own GGUF repository (XHToken/Spark-X2.6-4B-GGUF) ships Q4_K_M (3.54 GB), Q8_0 (4.38 GB) and an 8.23 GB F16 file. The spark2_5 architecture is registered in upstream llama.cpp b11847 and loads in Ollama 0.40.10 and 0.40.13 from an imported GGUF, but PrismML's llama.cpp fork at b10777 does not know it. No Ollama library tag exists (registry checked 2026-11-03). The vendor documents `ollama run SparkLLM/Spark-X2.6-4B`, a namespace tag that serves one 8.23 GB F16 file (manifest read 2026-11-03), which this catalog does not list: it is not a Q4_K_M, and that command would fetch the 8 GB file. The route tested here is to download the vendor's Q4_K_M and import it with `ollama create`. Measured on this project's own hardware from that Q4_K_M (docs/lab/radar-2026-10-30-model-validation.md), an RTX 3080 (10 GB) with nothing else on the card, all layers on the GPU: 170 tok/s and 7,438 tok/s prompt processing in upstream llama.cpp, 132 tok/s in Ollama; the server held 3,595 MiB at a 4,096 context, 4,631 MiB at 32,768 and 8,183 MiB at 131,072 (about 37 KiB of KV cache per token). On an RTX 4070 Ti desktop Ollama generated 106 to 109 tok/s and a 131,072 context peaked at 10,717 MiB. The model thinks at length by default: one coding run used a 12,000-token cap without reaching an answer, while with a 32,000-token cap all three thinking modes finished and passed (6,453 tokens by default, 426 with thinking off). The claimed 1,048,576-token context was tested only to 131,072. Ollama's reported VRAM size (api/ps) did not rise with context for this model, so budget from nvidia-smi rather than from it.