Leaderboard / SparkLLM (XHToken)

Spark-X2.6 4B

Name: Spark-X2.6 4B
Author: SparkLLM (XHToken)

SparkLLM (XHToken) · Spark-X2 · released 2026-10-19 · data checked 2026-11-01

Spark-X2.6 4B is a 4.1B-parameter dense model from SparkLLM (XHToken) (China), released 2026-10-19 under the Apache-2.0 license. It scores 67.5 on the Small LM Index, #2 of 21 tiny models. It needs about 3.4 GB of memory at 4-bit with an 8K-token context.

Thinking budget is configurable (none / low / high); scores use high. 1M-token context.

Small LM Index

67.5

#2 of 21 scored tiny models

Likely 65.0–69.9 · from 4 of 6 core benchmarks

Parameters

4.1B

Dense

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best tiny model on that benchmark. Sources: 11 reported by SparkLLM (XHToken). How we label sources.

GPQA Diamond Lab-reported Knowledge · GPQA Diamond77.1

Class best: Nanbeige4.2 3B 87.4

Humanity's Last Exam Lab-reported Reasoning17.9

AIME Lab-reported Math · AIME 202691.2

SWE-bench Verified Lab-reported Coding · Vendor harness62.8

Class best: Nanbeige4.2 3B 63.6

τ²-Bench Lab-reported Agentic86.9

Class best: MiniCPM5 2B 97.1

IFBench Lab-reported Instruction following75.8

IFEval Lab-reported Instruction following93.6

Class best: LFM3 1.7B 96.7

BFCL Lab-reported Agentic · BFCL v471.8

HMMT Lab-reported Math · HMMT Feb 202681.3

Class best: Nanbeige4.2 3B 82.8

AA-LCR Lab-reported Long context · AA-LCR v1.167.3

SciCode Lab-reported Coding36.8

Not reported: estimated

What the index model expects from this model's other scores. Not a measurement.

MMLU-Pro≈ 83

LiveCodeBench≈ 78

Specifications

Lab: SparkLLM (XHToken) (CN · China)
Architecture: Dense transformer
Total parameters: 4.1B
Active parameters: All (dense)
Size class: Tiny
Context window: 1M tokens
Input modalities: text
Reasoning: Hybrid (thinking on/off)
License: Apache-2.0
Openness: Open weights
Memory to run: ~3.4 GB at 4-bit, 8K context (more sizes)
Hugging Face: 18.9K downloads (30 days) · 977 likes

Weights on Hugging Face ↗

Compare with

Spark-X2.6 4B vs Nanbeige4.2 3B (4.2B, index 70.4)
Spark-X2.6 4B vs Spark-X2.5 4B (4.1B, index 65.5)
Spark-X2.6 4B vs Qwen3.5 4B (4B, index 57.8)
Spark-X2.6 4B vs Qwen3 4B Thinking 2507 (4B, index 55.7)
Spark-X2.6 4B vs Nemotron Labs Dory 1.0 4B (4B, index 60.0)
Spark-X2.6 4B vs Gemma 4 E2B (5.1B, index 31.9)

Ranked in

#9 in Best small open models for coding
#10 in Best small open models for math
#2 in Best LLMs that run on an 8 GB GPU or laptop
#2 in Best LLMs that run on a 16 GB GPU
#2 in Best tiny LLMs for phones and edge devices

Sources

Something wrong here? Report it.

Run it locally

Uses XHToken/Spark-X2.6-4B-GGUF (Q4_K_M). Swap the tag for another size: Q4_K_M, Q8_0.

Ollama

ollama run hf.co/XHToken/Spark-X2.6-4B-GGUF:Q4_K_M

llama.cpp

llama-server -hf XHToken/Spark-X2.6-4B-GGUF:Q4_K_M

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/Spark-X2.6-4B-OptiQ-4bit

vLLM (GPU server)

vllm serve XHToken/Spark-X2.6-4B

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	256K context
16-bit F16	9.0 GB	10 GB	19 GB
8-bit Q8_0	5.2 GB	6.2 GB	15 GB
4-bit Q4_K_M	3.3 GB	4.3 GB	13 GB

Check it against your GPU, Mac or phone →