Language modelOpen weights
Spark-X2.5-4B
iFLYTEK (SparkLLM Team)
- Released
- September 1, 2026
- Data date
- October 7, 2026
- Input
- Text
- Output
- Text
Spark-X2.5-4B uses hybrid attention, alternating one full-attention layer with three sliding-window layers. The open 4B edge checkpoint combines local attention with a native context window of 1,048,576 tokens.
Specifications and access
| Specification | Value and source |
|---|---|
| Parameters | 4B parametersSource |
| Context window | 1,048,576 tokens in the model configurationSource |
| Input modalities | TextSource |
| Output modalities | TextSource |
| Training stage | Post-trained checkpoint; the model card names large-scale reinforcement learning and MOPD.Source |
| License | Apache 2.0Source |
Page 1 of 2
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
BFCL v4 (Show measurement, test conditions, and source)
- Source value
- 65.1
- Score
- 65.1
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Benchmark version
- V4
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
τ²-bench (Show measurement, test conditions, and source)
- Source value
- 75.1
- Score
- 75.1
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
τ³-bench (Show measurement, test conditions, and source)
- Source value
- 30.4
- Score
- 30.4
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
MCP-Atlas (Show measurement, test conditions, and source)
- Source value
- 54.6
- Score
- 54.6
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
MCP-Mark (Show measurement, test conditions, and source)
- Source value
- 14.2
- Score
- 14.2
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
Workspace Bench (Show measurement, test conditions, and source)
- Source value
- 31.2
- Score
- 31.2
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
VitaBench 2.0 (Show measurement, test conditions, and source)
- Source value
- 25.2
- Score
- 25.2
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Benchmark version
- 2.0
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
BrowseComp (Show measurement, test conditions, and source)
- Source value
- 40.9
- Score
- 40.9
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
SWE-Bench Pro (Show measurement, test conditions, and source)
- Source value
- 44.4
- Score
- 44.4
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
SWE-Bench Verified (Show measurement, test conditions, and source)
- Source value
- 41.6
- Score
- 41.6
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
SWE-Bench Multilingual (Show measurement, test conditions, and source)
- Source value
- 53.3
- Score
- 53.3
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
SciCode (Show measurement, test conditions, and source)
- Source value
- 34.7
- Score
- 34.7
- Metric
- published score; the table does not state a scale or aggregation
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Official Spark-X2.5 model release repository
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official iFLYTEK table for this exact open edge checkpoint. The repository says all tests ran in thinking mode and recommends temperature 1.0, top_p 0.95 and top_k -1. Per-benchmark harnesses and sample counts are not published.
- Context
- The repository table does not state the measurement scale. The unchanged source value is therefore stored with an unknown unit.
Page 1 of 2
Sources and data date
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. Retrieved October 7, 2026. |
| Output | Documented modalities. Retrieved October 7, 2026. |
| Research status | Research date October 7, 2026. 2 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | score-scale-and-benchmark-setups Original wording The repository publishes scores but omits units, per-test harnesses, and sample counts for most benchmarks. Values remain unconverted and their units are unknown. |
| Additional source | Spark-X2.5-4B model configuration (retrieved October 7, 2026) · Spark-X2.5-4B model configuration · Editorial description reviewed October 7, 2026 |
| Additional source | Official Spark-X2.5 model release repository (retrieved October 7, 2026) · Official Spark-X2.5 model release repository · Editorial description reviewed October 7, 2026 |