AI Index

Qwen3-4B-Instruct-2507

Language Alibaba Qwen Partly verified Released Aug 6, 2025
  • Tool use
  • Open weights

Prices

0 offerings · prices as listed

No priced offering is listed for this model yet.

Details

Licence
apache-2.0
Family
qwen

Not stated on Alibaba Qwen’s pages: max input, input modalities, output modalities, reasoning, reasoning efforts, reasoning can be off, max thinking budget, structured output, built-in tools, knowledge cutoff, aliases.

Run locally

Weights: Qwen/Qwen3-4B-Instruct-2507
  • Parameters 4B
  • Dense
  • Native precision BF16
  • Layers 36
  • KV cache per token 144 KB
  • Native context 256K (model card)
  • Config positions 256K (config.json max_position_embeddings)
  • API context 256K (maker’s API)
My hardware GeForce RTX 4090 24 GB · 1008 GB/s · 64 GB DDR5 dual channel · 89.6 GB/s
On your hardware · GeForce RTX 4090
Runs well BF16 (native) (automatic) 8K context 54–82 tok/sOne GPU
WeightsKV cacheOverheadUsable
Needs
≈ 9.1 GB (weights 7.5 (maker’s file) + KV cache 1.1 + overhead 0.5)
Usable
21.6 GB (90% of its memory)
Generation speed
54–82 tok/s
Longest context
97K (longer doesn’t fit)
First token
about 0.13–0.19 s for a 2K prompt
PrecisionWeightsNeeds at 8KOn your hardwareLongest context
BF16 (native)maker’s file7.5 GB≈ 9.1Runs well 54–82 tok/s97K
Q8_04.0 GB≈ 5.6Runs well 92–138 tok/s122K
FP8maker’s file4.8 GB≈ 6.5Runs well 79–118 tok/s116K
Q6_K3.1 GB≈ 4.7Runs well 112–168 tok/s128K
Q5_K_M2.7 GB≈ 4.3Runs well 124–186 tok/s131K
Q5_02.6 GB≈ 4.2Runs well 127–190 tok/s131K
Q4_K_M2.3 GB≈ 3.9Runs well 138–207 tok/s133K
Q4_02.1 GB≈ 3.7Runs well 145–218 tok/s135K
IQ4_XS2.0 GB≈ 3.6Runs well 151–226 tok/s135K
Q3_K_Mquality drops1.8 GB≈ 3.5Runs well 159–238 tok/s137K
IQ3_XXSquality drops1.4 GB≈ 3.1Runs well 184–275 tok/s139K
Q2_Kquality drops1.4 GB≈ 3.0Runs well 186–278 tok/s140K
IQ2_XXSquality drops1.0 GB≈ 2.6Runs well 225–337 tok/s143K

Click a row to choose a precision. “Maker’s file” sizes are the maker’s own files; others are parameters × bits per weight.

Compared with other hardware

Each at the highest precision it runs. Click one for details.

Smallest that runs it (Q4_K_M, 8K): GPU GeForce RTX 3080; unified memory Apple M5 · 16 GB.

Run commands

vLLM
vllm serve Qwen/Qwen3-4B-Instruct-2507 --max-model-len 8192
Memory to fine-tune
Full≈ 62.1 GBDoesn’t fit your hardware
LoRA≈ 9.8 GBFits your hardware
QLoRA≈ 4.5 GBFits your hardware

Batch 1, 2,048-token sequences, gradient checkpointing; LoRA and QLoRA train rank-16 adapters on the attention projections.

Estimates; facts from the maker’s Hugging Face repositories, as of 2026-11-06. How it’s estimated

Cited scores

From third-party leaderboards, not measured by us
  • Epoch AI · GPQA Diamond 45.8%

    #139 of 168 listed models · ±2.85

  • Epoch AI · OTIS Mock AIME 2024–2025 52.2%

    #110 of 159 listed models · ±6.59

  • Epoch AI, 'Capabilities & benchmarking'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/benchmarks'. Licensed CC BY 4.0. · License · as of Nov 4, 2026
  • changes we made

Changes

All changes →

No changes recorded yet.

Sources

Alibaba Qwen’s official policies: data use, privacy, terms →

Some values on this page are still waiting to be checked against official pages: a filled dot is verified, a hollow one is not yet (with or without a source on file). How we verify data

Alibaba Qwen · China · qwen.ai

Snapshots, the same family and other models from this provider