Leaderboard / Alibaba (Qwen)

Qwen3.5 2B

Name: Qwen3.5 2B
Author: Alibaba (Qwen)

Alibaba (Qwen) · Qwen3.5 · released 2026-03-02 · data checked 2026-11-02

Qwen3.5 2B is a 2.3B-parameter dense model from Alibaba (Qwen) (China), released 2026-03-02 under the Apache-2.0 license. It scores 33.0 on the Small LM Index, #14 of 21 tiny models. It needs about 2 GB of memory at 4-bit with an 8K-token context.

Small LM Index

33.0

#14 of 21 scored tiny models

Likely 28.8–37.4 · from 2 of 6 core benchmarks

Parameters

2.3B

Dense

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

262K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best tiny model on that benchmark. Sources: 10 reported by Alibaba (Qwen). How we label sources.

MMLU-Pro Lab-reported Knowledge · Thinking mode (all lab scores)66.5

Class best: Qwen3.5 4B 79.1

GPQA Diamond Lab-reported Knowledge51.6

Class best: Nanbeige4.2 3B 87.4

τ²-Bench Lab-reported Agentic48.8

Class best: MiniCPM5 2B 97.1

IFBench Lab-reported Instruction following41.3

Class best: Spark-X2.6 4B 75.8

MMLU Lab-reported Knowledge · MMLU-Redux79.6

Class best: MiniCPM5 2B 84.7

Multilingual MMLU Lab-reported Multilingual63.1

Class best: SmolLM3 3B 64.1

IFEval Lab-reported Instruction following78.6

Class best: LFM3 1.7B 96.7

BFCL Lab-reported Agentic · BFCL v443.6

Class best: Spark-X2.6 4B 71.8

HMMT Lab-reported Math · HMMT Feb 202522.9

Class best: Nanbeige4.2 3B 82.8

AA-LCR Lab-reported Long context25.6

Class best: Spark-X2.6 4B 67.3

Not reported: estimated

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

Humanity's Last Exam≈ 3

AIME≈ 39

LiveCodeBench≈ 36

SWE-bench Verified≈ 14

Specifications

Lab: Alibaba (Qwen) (CN · China)
Architecture: Dense transformer
Total parameters: 2.3B
Active parameters: All (dense)
Size class: Tiny
Context window: 262K tokens
Input modalities: text, image, video
Reasoning: Hybrid (thinking on/off)
License: Apache-2.0
Openness: Open weights
Memory to run: ~2 GB at 4-bit, 8K context (more sizes)
Hugging Face: 3.4M downloads (30 days) · 465 likes

Weights on Hugging Face ↗

Compare with

Qwen3.5 2B vs LFM3 1.7B (1.7B, index 60.0)
Qwen3.5 2B vs MiniCPM5 2B (2.5B, index 57.5)
Qwen3.5 2B vs Spark-X2.6 1.7B (1.7B, index 50.1)
Qwen3.5 2B vs LFM2.5 2.6B (2.7B, index 43.2)
Qwen3.5 2B vs Granite 4.2 3B (3B, index 54.8)
Qwen3.5 2B vs Ministral 3 3B (3B, index 49.4)

Sources

Something wrong here? Report it.

Run it locally

Uses unsloth/Qwen3.5-2B-GGUF (Q4_K_M). Swap the tag for another size: BF16, IQ4_NL, IQ4_XS, Q3_K_M, Q3_K_S, Q4_0, Q4_1, Q4_K_M….

Ollama

ollama run hf.co/unsloth/Qwen3.5-2B-GGUF:Q4_K_M

llama.cpp

llama-server -hf unsloth/Qwen3.5-2B-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/Qwen3.5-2B-MLX-4bit

Text only; for images use mlx-vlm.

vLLM (GPU server)

vllm serve Qwen/Qwen3.5-2B

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	5.2 GB	5.5 GB	8.4 GB
8-bit Q8_0	3.0 GB	3.4 GB	6.2 GB
4-bit Q4_K_M	2.0 GB	2.3 GB	5.2 GB

Check it against your GPU, Mac or phone →