Model
Model explorer

Spark-X2.6-4B

OPEN
iFlytek · Spark (Xinghuo) family · released Oct 20, 2026

iFlytek's open-weight on-device model, released by its SparkLLM (XHToken) team on Hugging Face on Oct 20, 2026 (the GitHub release log dates the global launch Oct 21) and presented again at iFlytek's 1024 Developer Festival on Oct 22 alongside the cloud Spark X2.6. It keeps Spark-X2.5-4B's architecture and gains through post-training; a base checkpoint ships separately. Scores are SparkLLM's with thinking set to high. NOT FILED: BFCL-V4 (71.8; corpus has only bfcl and bfcl-v3, and V4 is a different task version) and AA-LCR v1.1 (67.3; no corpus column). Spark-X2.5-4B (Sep 1, 2026) predates this routine and is not in the corpus.

ReasoningCodingVisionFunction callingTool useAgentic
1760.2
Elo · rank #119
Parameters
4B
Active params
4B (dense)
Context
1.024M tokens
Architecture
4B-parameter on-device language model with the same architecture as Spark-X2.5-4B (Spark2_5ForCausalLM; layer layout not described on the card); 1,048,576-token context; thinking levels none, low and high
License
Apache-2.0
Languages
—
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
BrowseCompAgents53.1%#37
best: Kimi K3.1 Pro · 92.1%
MCP AtlasAgents62.4%#9
best: Muse Spark Open · 89.2%
SWE-bench ProCoding54.3%#41
best: Claude Opus 5.5 · 89.9%
SWE-bench VerifiedCoding62.8%#96
best: Claude Opus 5 · 97.0%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
—
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
GGUF
Fine-tune it
Permissive
QLoRA4.5 GB1× RTX 3060 12GB
LoRA10.3 GB1× RTX 3060 12GB
Full fine-tune66.0 GB1× A100 80GB
QLoRA SFT on ~10k samples ≈ $2.34 (1× RTX 3060 12GB)
API price weights · each benchmark row carries its own source badge (see methodology)