1.708B dense (1,707,657,216)
1,048,576 tokens
Apache-2.0
Second model in the Spark-X2.6 series and a distinct later release from Spark-X2.6-4B. The compact English/Chinese text model retains exactly the same architecture as Spark-X2.5-1.7B and attributes its capability gains to improved post-training rather than architectural scaling. Its 28-layer dense Transformer repeats three 512-token sliding-attention layers followed by one full-attention layer, producing seven full-attention layers across the network. The backbone uses a 2,048-dimensional hidden state, 6,656-dimensional feed-forward layers, eight query heads, two key/value heads, 256-dimensional attention heads, tied input/output embeddings, a 131,072-token vocabulary, and headwise sigmoid attention-output gates. Full-attention layers use 25% partial rotary embeddings with a 5,000,000 RoPE base, while sliding layers use full rotary embeddings with a 10,000 base. The native configuration supports 1,048,576 tokens. Its chat template supports structured tools, explicit reasoning content, thinking disabled or enabled, and low or high thinking effort, with high enabled by default. The publisher describes improvements in coding, complex tool use, multi-step planning, software engineering, multilingual behavior, reasoning, and long-context work, but does not disclose the new post-training recipe, data, token budget, or compute. In high-thinking publisher evaluations, Spark-X2.6-1.7B reports a 49.6 average across 30 benchmarks versus 35.8 for Spark-X2.5-1.7B, 22.0 for Qwen3.5-2B, 22.5 for Gemma4-E2B, 46.9 for MiniCPM5-2B, and 30.4 for LFM2.5-2.6B; Spark-X2.6-4B scores 63.3 as a larger reference. Individual reported results include 55.2 on BFCL-V4, 85.3 on tau2-bench, 41.8 on MCP-Atlas, 66.9 on ClawEval-EN, 41.7 on BrowseComp, 58.8 on PAgentBench, 59.8 on SWE-bench Verified, 57.3 on SWE-bench Multilingual, 70.1 on AIME 2026, 65.5 on IFBench, 89.4 on IFEval, and 56.1 on GPQA Diamond. Long-context results include 85.1 on Graphwalks BFS at 256K, 25.7 at 1M, 56.1 on OpenAI MRCR v2 from 256K to 512K, and 19.9 from 512K to 1M. All evaluations use thinking mode with Spark set to high; some comparison-model results marked by the publisher are copied from public model cards or papers, while SWebBench and PAgentBench remain internal benchmarks. The official BF16 checkpoint contains exactly 1,707,657,216 parameters across two safetensor shards totaling approximately 3.42 GB. Official releases also include the base model, FP8, INT8, INT4, and GGUF variants, with deployment guidance for Transformers, SGLang, vLLM, llama.cpp, NVIDIA GPUs, and Huawei Ascend NPUs. XHToken reports leading short-context decoding throughput against Qwen3.5-2B, MiniCPM5-2B, and LFM2.5-2.6B in its Apple M5 Pro and RTX 4090 tests, with an additional optimized llama.cpp fork improving the RTX results.