LOADING DATASHEET
GROUNDED IN REAL HUMAN CONVERSATIONSYOUR BENCH
Qwen3 VL 4B Instruct
◆ OPEN
by Alibaba
OPEN WEIGHTSLOCALCHINA-MADEAGENTSSMALL
RANKED #290 OF 324 ASSISTANTS · OVERALL #454 OF 7,062 · VIBE SCORE 5.1 · 25 VOICES
Huihui-Qwen3-VL-4B-Instruct-abliterated-GGUF on Hugging Face (image text to text). 8,947 downloads. Open weights for local or hosted use.
HAPPY6.3
2 mentions
SAD8.7
3 mentions
VIBE TRAJECTORY
26 weeks · 23 voices
HOLDING UP
Recent vibe is steady. No clear fade in the crowd.
HOW TO USE IT
Open weights. You can use a hosted service, or download it and run it yourself, free.
WHAT THE INTERNET SAYS
People are mostly frustrated with Qwen3 VL 4B Instruct right now.
19 POSTS · 6 COMMENTS · REDDIT · GITHUB
PEOPLE LOVE
- Stable and dependable (1 mentions)
PEOPLE COMPLAIN ABOUT
- Outages, errors, or limits get in the way (6 mentions)
TOPIC BY TOPIC
- Working when you need itA SORE SPOT
1 thumbs up · 6 thumbs down
LOOKING FOR AN ALTERNATIVE?
Other models the crowd has fully reviewed, starting with multimodal models like this one.
Grok 4.6> SOURCES (12)
- REDDIT · 2026-01-25 · 385 UPVOTESOllama Models Ranked by VRAM Requirements
Ollama Models Ranked by VRAM Requirements
- REDDIT · 2026-07-28 · 76 UPVOTESmicrosoft/Mage-VL · Hugging Face - An Efficient Codec-Native Streaming Multimodal Foundation Model
microsoft/Mage-VL · Hugging Face - An Efficient Codec-Native Streaming Multimodal Foundation Model
- REDDIT · 2026-07-29 · 67 UPVOTESKrea 2 LoRA training on a 16GB RTX 5080: full measurements, and four sourced corrections to the guidance going around
Krea 2 LoRA training on a 16GB RTX 5080: full measurements, and four sourced corrections to the guidance going around
- REDDIT · 2026-06-30 · 63 UPVOTESKrea2 GGUF Worklfow for 8-12GB VRAM
Krea2 GGUF Worklfow for 8-12GB VRAM
- REDDIT · 2026-07-27 · 62 UPVOTESTrying out LoKr instead of LoRA on Krea2
Trying out LoKr instead of LoRA on Krea2
- REDDIT · 2025-12-22 · 2 UPVOTESWhat do you use for image-to-text? This one doesn't seem to work
What do you use for image-to-text? This one doesn't seem to work
- GITHUB · 2026-09-21 · 1 UPVOTESWrite skip-module names vLLM can match after it fuses the MLP projections
Write skip-module names vLLM can match after it fuses the MLP projections
- GITHUB · 2026-08-25 · 0 UPVOTESAttempting to generate FLUX.2 Klein images from inside Canvas reports "no Z-Image model available"
Attempting to generate FLUX.2 Klein images from inside Canvas reports "no Z-Image model available"
- GITHUB · 2026-08-25 · 0 UPVOTESDeployed FLUX.2 Klein LoRAs always report as "not deployed" in Canvas
Deployed FLUX.2 Klein LoRAs always report as "not deployed" in Canvas
- GITHUB · 2026-09-08 · 0 UPVOTESperf(qwen_vl): multimodal RoPE leaks f32 into the decode residual
perf(qwen_vl): multimodal RoPE leaks f32 into the decode residual
- GITHUB · 2026-09-15 · 0 UPVOTES[BUG]: Agent mode (@agent / Agent chat mode) never returns a response with local Ollama provider — request aborted immediately (AbortError)
[BUG]: Agent mode (@agent / Agent chat mode) never returns a response with local Ollama provider — request aborted immediately (AbortError)
- GITHUB · 2026-09-22 · 0 UPVOTESllama_cpp backend: VLM vision output silently wrong on compute_unit=npu (HTP0), crashes on compute_unit=gpu
Describe the bug Running a Qwen3-VL-architecture GGUF VLM via runtime_id=llama_cpp with an image input produces different, backend-dependent results: - compute_unit=npu (HTP0/Hexagon): generation completes with no error, but the vision output is silently…