Leaderboard / Alibaba (Qwen)

Qwen3.5 0.8B

Name: Qwen3.5 0.8B
Author: Alibaba (Qwen)

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

Qwen3.5 0.8B is a 900M-parameter dense model from Alibaba (Qwen) (China), released 2026-03-02 under the Apache-2.0 license. It scores 10.7 on the Small LM Index, #20 of 21 tiny models, with 1 independently measured benchmark. It needs about 1.2 GB of memory at 4-bit with an 8K-token context.

Small LM Index

10.7

#20 of 21 scored tiny models

Likely 8.9–13.0 · from 2 of 6 core benchmarks

Parameters

900M

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: 9 reported by Alibaba (Qwen), 1 independently measured. How we label sources.

MMLU-Pro Lab-reported Knowledge · Thinking mode (all lab scores)42.3

Class best: Qwen3.5 4B 79.1

GPQA Diamond Lab-reported Knowledge11.9

Class best: Nanbeige4.2 3B 87.4

τ²-Bench Lab-reported Agentic11.6

Class best: MiniCPM5 2B 97.1

IFBench Lab-reported Instruction following21.0

Class best: Spark-X2.6 4B 75.8

MMMU-Pro Independent Vision · Measured by Artificial Analysis25.8

Class best: Qwen3.5 4B 65.4

MMLU Lab-reported Knowledge · MMLU-Redux59.5

Class best: MiniCPM5 2B 84.7

Multilingual MMLU Lab-reported Multilingual44.3

Class best: SmolLM3 3B 64.1

IFEval Lab-reported Instruction following44.0

Class best: LFM3 1.7B 96.7

BFCL Lab-reported Agentic · BFCL v425.3

Class best: Spark-X2.6 4B 71.8

AA-LCR Lab-reported Long context4.7

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≈ 0

AIME≈ 4

LiveCodeBench≈ 9

SWE-bench Verified≈ 2

Specifications

Lab: Alibaba (Qwen) (CN · China)
Architecture: Dense transformer
Total parameters: 900M
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: ~1.2 GB at 4-bit, 8K context (more sizes)
Hugging Face: 3.7M downloads (30 days) · 801 likes

Weights on Hugging Face ↗

Compare with

Qwen3.5 0.8B vs LFM2.5 1.2B Thinking (1.2B, index 30.1)
Qwen3.5 0.8B vs LFM2.5 1.2B Instruct (1.2B, index 20.6)
Qwen3.5 0.8B vs LFM2.5 350M (350M, index 12.6)
Qwen3.5 0.8B vs Gemma 3 1B (1B, index 6.1)
Qwen3.5 0.8B vs EXAONE 4.0 1.2B (1.3B, index 36.1)

Sources

Something wrong here? Report it.

Run it locally

Uses ggml-org/Qwen3.5-0.8B-GGUF (Q4_0). Swap the tag for another size: BF16, Q4_0, Q8_0.

Ollama

ollama run hf.co/ggml-org/Qwen3.5-0.8B-GGUF:Q4_0

llama.cpp

llama-server -hf ggml-org/Qwen3.5-0.8B-GGUF:Q4_0

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

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/Qwen3.5-0.8B-MLX-4bit

Text only; for images use mlx-vlm.

vLLM (GPU server)

vllm serve Qwen/Qwen3.5-0.8B

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	2.4 GB	2.7 GB	5.5 GB
8-bit Q8_0	1.5 GB	1.9 GB	4.7 GB
4-bit Q4_K_M	1.1 GB	1.5 GB	4.3 GB

Check it against your GPU, Mac or phone →