Model Details
Spark-X2.6-4B
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 | 62.8 | XHToken HF model card | 2026-10 |
| GPQA Diamond | 77.1 | XHToken HF model card | 2026-10 |
| AIME 2026 | 91.2 | XHToken HF model card | 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 | 1M | 41.6 | 73.7 | Open weights |
| 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
- Benchmarks: SWE-bench Verified · GPQA Diamond · 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.