# Spark-X2.6-4B ![Organization](https://img.shields.io/badge/Org-XHToken-555) ![Context](https://img.shields.io/badge/Context-1M%20tokens-1f6feb) ![Pricing](https://img.shields.io/badge/Price-Open%20weights-2da44e) ![License](https://img.shields.io/badge/License-Apache%202.0-d29922) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) ![Released](https://img.shields.io/badge/Released-Oct%202026-555) > **Quick answer:** Spark-X2.6-4B is XHToken's refreshed compact general-purpose language model, retaining Spark-X2.5-4B's hybrid-attention architecture (one full-attention layer plus three sliding-window layers, native 1,048,576-token context) while gaining substantial post-training improvements in coding, agentic workflows, long-context, multilingual, and reasoning tasks. Open weights, Apache 2.0. ## At a Glance **Where Spark-X2.6-4B leads** - Strong agentic tool-use scores for its size class: BFCL-V4 71.8, MCP-Atlas 62.4, BrowseComp 53.1 - Solid coding performance: SWE-bench Verified 62.8, SWE-bench Pro 54.3 - Controllable thinking budget (none/low/high) to trade reasoning quality for latency/cost **Where it lags** - Still trails larger ~9B-class models on some reasoning benchmarks (e.g. GPQA Diamond 77.1 vs some rivals) - English/Chinese-focused; narrower language coverage than some peers **Best for:** on-device/edge agentic deployments needing strong tool-use and coding at a 4B footprint. ## What Spark-X2.6-4B Is Spark-X2.6-4B is SparkLLM/XHToken's October 2026 refresh of their compact agentic model line. It keeps the exact architecture of Spark-X2.5-4B — a hybrid-attention design combining one full-attention layer with three sliding-window attention layers, enabling a native 1M-token context window without the typical compute overhead of long-context dense models. The gains over X2.5 come entirely from improved post-training rather than architecture changes, with the vendor reporting state-of-the-art results among similarly sized open-source models on agentic, coding, and long-context tasks — in some cases matching or exceeding ~9B-class models. ## Specifications | Field | Value | |---|---| | Organization | XHToken (SparkLLM) | | Parameters | 4B (dense) | | Context window | 1,048,576 tokens | | Architecture | Hybrid attention (1 full-attention + 3 sliding-window layers per block) | | License | Apache 2.0 | | Release date | 2026-10-21 | | Modality | Text | ## Pricing | | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | Open weights | Free to download | Free to download | Open weights — self-hosting cost only. No official hosted API pricing published at time of writing. ## Context Window Spark-X2.6-4B has a **1,048,576-token (1M) context window** — enough for entire codebases or very long multi-document analysis in a single request, enabled by its hybrid sliding-window/full-attention design that keeps KV-cache overhead manageable at this scale. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [SWE-bench Verified](/benchmarks/princeton-nlp-openai/swe-bench-verified) | 62.8 | [XHToken HF model card](https://huggingface.co/XHToken/Spark-X2.6-4B) | 2026-10 | | [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) | 77.1 | [XHToken HF model card](https://huggingface.co/XHToken/Spark-X2.6-4B) | 2026-10 | | [AIME 2026](/benchmarks/mathematical-association-of-america/aime-2026) | 91.2 | [XHToken HF model card](https://huggingface.co/XHToken/Spark-X2.6-4B) | 2026-10 | *Scores above are reported by XHToken and shown for context. They are not Benchgen measurements. Figures depend on harness, tools, and effort settings — see the source for methodology.* ## Spark-X2.6-4B vs Alternatives | Model | Context | SWE-bench Verified | GPQA Diamond | Price (in/out per 1M) | |---|---|---|---|---| | **Spark-X2.6-4B** | 1M | 62.8 | 77.1 | Open weights | | [Spark-X2.5-4B](/models/xhtoken/spark-x2-5-4b) | 1M | 41.6 | 73.7 | Open weights | | [Qwen3.5-4B](/models/alibaba/qwen3-5-4b) | — | 38.8* | 67.2 | Open weights | Spark-X2.6-4B is a post-training-only refresh of X2.5, and the jump is large on coding and agent tasks specifically — a builder already on X2.5 gets meaningfully better SWE-bench/agentic performance for the same deployment footprint. ## How Spark-X2.6-4B Performs on Real Agent Tasks The model's BFCL-V4, MCP-Atlas, and BrowseComp scores point to competent function-calling, multi-tool orchestration, and web-research agentic loops at a 4B parameter budget. As with any vendor-reported benchmark, these numbers reflect curated harnesses and don't guarantee reliability on a specific production tool stack — treat them as a capability signal, not a deployment guarantee. ## Related Models & Benchmarks - Compare: [Spark-X2.5-4B](/models/xhtoken/spark-x2-5-4b) - Benchmarks: [SWE-bench Verified](/benchmarks/princeton-nlp-openai/swe-bench-verified) · [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) · [AIME 2026](/benchmarks/mathematical-association-of-america/aime-2026) ## Frequently Asked Questions

What is Spark-X2.6-4B?

A 4B-parameter open-weight language model from XHToken (SparkLLM), a post-training refresh of Spark-X2.5-4B with stronger coding, agent, and long-context scores.

What is Spark-X2.6-4B's context window?

1,048,576 tokens (1M) natively, via a hybrid full-attention/sliding-window architecture.

How much does Spark-X2.6-4B cost?

Open weights — free to download, hosting cost only.

Is Spark-X2.6-4B open source?

Yes, released under the Apache 2.0 license.

What is Spark-X2.6-4B's knowledge cutoff?

Not publicly disclosed at time of writing.

--- **Excluded vendor benchmarks (bucket C, not published to leaderboard — documented here per Scout's data-ops rules):** - τ²-bench, τ³-bench (generic, unscoped — no domain qualifier given, could map to multiple existing domain-specific entities; excluded rather than guessed) - VitaBench2.0, Gaokao 2026, IMO-AnswerBench, Graphwalks bfs 256k/1M, Openai_MRCR v2 256k-512k/512k-1M — no matching local benchmark entity (either no entity exists, or only a different-length/version variant exists, which would be a methodology mismatch) - EnterpriseOps-Gym, ClawEval-EN, SWebBench, PAgentBench — vendor-internal/in-house evals explicitly noted as "will be open-sourced soon," no independent reproducibility yet - Plain "GPQA" (69.9, full set) — excluded separately from "GPQA Diamond" (77.1, published above) since they are different question subsets; only Diamond maps cleanly to our existing entity *Specs and scores sourced from XHToken's official HF model card; third-party benchmark scores attributed inline. Last updated 2026-10-23.*

Spark X2.6 4B — Evaluations

Evaluations