LOADING DATASHEET
GROUNDED IN REAL HUMAN CONVERSATIONSYOUR BENCH
Qwen3 Moe Tiny
◆ OPEN
by Alibaba
OPEN WEIGHTSLOCALCHINA-MADE
RANKED #266 OF 318 ASSISTANTS · OVERALL #407 OF 494 RANKED · VIBE SCORE 5.4 · 54 VOICES ·
COPY BADGE pastes a README snippet with the rank SVG.
qwen3-moe-tiny on Hugging Face (text generation). 2,149 downloads. Open weights for local or hosted use.
HAPPY8.6
2 mentions
SAD6.4
1 mentions
VIBE TRAJECTORY
26 weeks · 55 voices in this window
HONEYMOON FADING
Early praise is cooling off in recent weeks.
HOW TO USE IT
Open weights. You can use a hosted service, or download it and run it yourself, free.
RUN IT ON YOUR OWN MACHINE
Open weights: download Qwen3 Moe Tiny once and it's yours. No subscription, no rate limits, works offline.
670M parameters (MoE architecture), a Q4 quantized file is about 450 MB while unquantized FP16 is around 1.4 GB
- BARE MINIMUM
Raspberry Pi 5 (8GB)
8 GB RAM
~$80 NEWQ4 quantization, ~35 tok/s, 8k context via CPU inference
- SWEET SPOT
Apple Mac mini M2 (8GB)
8 GB UNIFIED RAM
~$450 USEDFP16 full precision, ~120 tok/s, 32k context on Metal GPU
- FULL POWER
Nvidia GeForce RTX 3060 (12GB)
12 GB VRAM
~$220 USEDFP16 full precision, ~220 tok/s, 32k context with high batch throughput
LEARN HOW TO SET IT UP
WILL MY BOX RUN QWEN3 MOE TINY?
The three cards above are researched picks. Search the machine you actually have — we only say yes if it should reply at a usable speed, not just load the weights and crawl.
HARDWARE PICKS & PRICES ARE RESEARCHED FROM THE LIVE WEB AND REFRESHED AUTOMATICALLY. TREAT THEM AS BALLPARK, NOT GOSPEL.
WHAT CAN YOU ACTUALLY DO WITH IT?
Everyday jobs on each of those rigs: how long Qwen3 Moe Tiny takes, and how good the crowd says it is at that kind of work.
| THE JOB | BARE MINIMUMRaspberry Pi 5 (8GB) | SWEET SPOTApple Mac mini M2 (8GB) | FULL POWERNvidia GeForce RTX 3060 (12GB) |
|---|---|---|---|
| Write an emaila paragraph or two, drafted from a one-line brief | ~14 S | ~4 S | ~2 S |
| Summarize a documenta long report boiled down to the points that matter | ~43 S | ~13 S | ~7 S |
| Build a websitea small landing page, markup and styles together | ~3.8 MIN★★★★★ | ~67 S★★★★★ | ~36 S★★★★★ |
| Build a backendan API with routes, storage and tests | ~12 MIN★★★★★ | ~3.5 MIN★★★★★ | ~1.9 MIN★★★★★ |
| Build a gamea playable browser game in one file | ~7.1 MIN★★★★★ | ~2.1 MIN★★★★★ | ~68 S★★★★★ |
STARS ARE THE COMMUNITY SCORE FOR THAT KIND OF WORK, DOCKED FOR HOW SQUASHED THE WEIGHTS GET AT EACH TIER. TIMES ARE COMPUTED FROM RESEARCHED SPEEDS. BALLPARK, NOT GOSPEL.
WHAT THE INTERNET SAYS
Not enough discussion about Qwen3 Moe Tiny yet to call it. We found 54 posts but people did not say much either way.
52 POSTS · 2 COMMENTS · GITHUB · REDDIT
PEOPLE LOVE
- No fan mail yet. The crowd is still warming up.
PEOPLE COMPLAIN ABOUT MORE
- No complaints yet. The internet is still looking for one. Suspicious.
LOOKING FOR AN ALTERNATIVE? SEE ALL
Other models the crowd has fully reviewed, starting with text models like this one.
Muse Spark 1.3 Contributor
DeepSeek V4 Flash> SOURCES (12)
- GITHUB · 2026-09-10 · 1 UPVOTESfix(container/vllm): patch exaone_moe lm_head prefix + bump vLLM to v0.28.0 for K-EXAONE
fix(container/vllm): patch exaone_moe lm_head prefix + bump vLLM to v0.28.0 for K-EXAONE
- GITHUB · 2026-08-27 · 1 UPVOTESfeat(mlx): support Qwen3.8-Flash-Next and GLM-5.3-Flash — differentiable MoE routing, structural patch selection, and text-path loading
feat(mlx): support Qwen3.8-Flash-Next and GLM-5.3-Flash — differentiable MoE routing, structural patch selection, and text-path loading
- GITHUB · 2026-09-10 · 1 UPVOTESfix(container/vllm): patch exaone_moe lm_head prefix so K-EXAONE 2.0 NVFP4 loads
fix(container/vllm): patch exaone_moe lm_head prefix so K-EXAONE 2.0 NVFP4 loads
- GITHUB · 2026-10-11 · 1 UPVOTESload_in_fp8 + fast_inference: train on vLLM's native online FP8 instead of torchao
load_in_fp8 + fast_inference: train on vLLM's native online FP8 instead of torchao
- GITHUB · 2026-09-17 · 1 UPVOTES[Test][MoE] Restore v0.27.1 external Qwen gate for acceptance control
[Test][MoE] Restore v0.27.1 external Qwen gate for acceptance control
- GITHUB · 2026-09-07 · 0 UPVOTESAlign tiny/small Qwen2.5 config with Qwen/Qwen2.5-32B-Instruct
Align tiny/small Qwen2.5 config with Qwen/Qwen2.5-32B-Instruct
- GITHUB · 2026-09-08 · 0 UPVOTESFix qwen3_moe.py sanitize() gating and update DeepSeek-V3.2 to use PipelineMixin
Fix qwen3_moe.py sanitize() gating and update DeepSeek-V3.2 to use PipelineMixin
- GITHUB · 2026-09-22 · 0 UPVOTES[rollout, vllm] fix: run the standard weight sync through vLLM's layerwise reload lifecycle
[rollout, vllm] fix: run the standard weight sync through vLLM's layerwise reload lifecycle
- GITHUB · 2026-08-16 · 0 UPVOTESfix(ENG-EXPERT-STREAM): the step clock had no caller, so the decode number measured a dead cache (#912, #1066)
fix(ENG-EXPERT-STREAM): the step clock had no caller, so the decode number measured a dead cache (#912, #1066)
- GITHUB · 2026-08-16 · 0 UPVOTESfix(ENG-EXPERT-STREAM): the step clock had no caller, so the decode number measured a dead cache (#912, #1066)
fix(ENG-EXPERT-STREAM): the step clock had no caller, so the decode number measured a dead cache (#912, #1066)
- GITHUB · 2026-08-16 · 0 UPVOTESfix(ENG-EXPERT-STREAM): the step clock had no caller, so the decode number measured a dead cache (#912, #1066)
fix(ENG-EXPERT-STREAM): the step clock had no caller, so the decode number measured a dead cache (#912, #1066)
- GITHUB · 2026-08-18 · 0 UPVOTES[AMD][CI] Prune dead Qwen coverage from the AMD nightly
[AMD][CI] Prune dead Qwen coverage from the AMD nightly