Now available — Qwen3.7-30B-Open weights on Hugging Face MIT licensed · No gated access · No waitlist 30B parameters · 262K context · 119 languages vLLM · SGLang · llama.cpp · MLX ready Runs on a single 3090 (Q4_K_M) 📜 Paper: arxiv:2603.14217 🪶 Probably the last 30B you'll ever need. Now available — Qwen3.7-30B-Open weights on Hugging Face MIT licensed · No gated access · No waitlist 30B parameters · 262K context · 119 languages vLLM · SGLang · llama.cpp · MLX ready Runs on a single 3090 (Q4_K_M) 📜 Paper: arxiv:2603.14217 🪶 Probably the last 30B you'll ever need.
Qwen 3 Series · Open Weights · v3.7.0

Qwen3.7-30B is
finally here.

The community can finally stop refreshing Hugging Face every 11 seconds. The weights are out, the license is real, and the autorefresh extension has been uninstalled. You're welcome.

✓ MIT License
⚖ No gating
🪶 Apache 2.0 compatible
🛡 OpenRAIL-M
🏠 Local inference
Qwen 3 · 30B Base
Open
30B
parameters
Dense · 64 layers · GQA · RoPE θ=1M
Context length
262K tokens
Open weights
Long context
Tool use
Coding
Reasoning
Local inference
0
Parameters
0
Context tokens
0
Languages
0
MMLU-Pro
0
HumanEval+
0
Open weights

What it can do, in a single model

30B parameters, no asterisks, no "but wait until the next version". Ship-ready across reasoning, code, agents, and 119 languages.

Reasoning that doesn't lie

Chain-of-thought, tool calling, and step-by-step planning that you can actually audit. No more "as an AI model" cop-outs.

Coding & agentic workflows

Strong on HumanEval+, LiveCodeBench, SWE-bench Verified. Edit files, run commands, and ship PRs before your coffee gets cold.

262K token context

Drop in a whole codebase, a book, or your last 6 months of chat history. It will remember more than you do.

Tool use & function calling

Native function calling, JSON-mode, and structured outputs. Plays nicely with LangChain, LlamaIndex, and your own spaghetti.

Multilingual, properly

119 languages with strong performance, not just "passes the hello world test". Including Klingon*.

Runs on your desk

Quantized GGUF fits in 18GB. A single 3090 is enough. Two 3090s in NVLink is glorious. The cloud is optional.

Beating everyone, including yourself

Independent evals, no cherry-picking, and one very honest row about your current coping mechanism.

MMLU-Pro — 14 disciplines, no multiple choice lifelines
/ 100
GPQA Diamond — PhD-level science, no googling
/ 100
ARC-AGI (v2) — Abstraction & reasoning, the hard one
/ 100
HumanEval+ — 164 problems, strict tests
/ 100
LiveCodeBench — Real LeetCode, dated, no contamination
/ 100
SWE-bench Verified — Fix real GitHub issues
/ 100
MATH Olympiad — 850 problems, olympiad-level
/ 100
AIME 2025 — High school math olympiad
/ 100
GSM8K — Grade school math, for the vibes
/ 100
BBH (Big Bench Hard) — 23 hard reasoning tasks
/ 100
HLE (Humanity's Last Exam) — The one nobody wants to talk about
/ 100
MMLU-Redux — Cleaned-up MMLU, no duplicates
/ 100

Models compared: Qwen3.7-30B-Open (us) · GPT-7.2 Orion · Claude 5.1 Sonnet-Opus · Llama 4 70B Closed · Your current coping mechanism

Pick your flavor of openness

Pick a format, smash the button, and go. Checksums below. Yes, the torrent is real.

60.2 GBFP16 · safetensors

Hugging Face

The full BF16 weights. Use this if you have the VRAM, the patience, or both.

safetensorssharded 4/4config.json
sha256: a7c2...f9d1e4b8c0...3a7d2e9f...b1c4d6e8...f0a2b3c4
17.2 GBQ4_K_M · GGUF

GGUF (llama.cpp)

Fits on a single 3090. Runs on a Mac mini. The everyman's choice.

Q4_K_MQ5_K_MQ8_0llama.cpp
sha256: b9d4...7e2c1a8f0...5d3b9c2e...8a4f1d6b...c7e2a3b5
$0 / moUp to 1M tokens/day

OpenRouter API

If you don't want to deal with VRAM thermals, just hit the API. OpenAI-compatible.

OpenAI compatstreamingtools
endpoint: https://api.qwen3-7-open.fake/v1
17.2 GB· 1,247 seeders

Torrent

For when the Hugging Face mirror is having a moment. Magnet link below.

magnetqBittorrentdecentralized
magnet: ?xt=urn:btih:a1b2c3d4e5f6...&tr=udp://tracker.openbittorrent.com:80

A brief history of "soon™"

For the historians, the cope-posters, and everyone who set an autorefresh cron job at 2am.

Jan 2024
Training begins
A reasonable amount of compute. A reasonable amount of hope.
Mar 2024
"Qwen 3 is coming soon™"
Posted from a fresh account. The replies were a mistake.
Jun 2024
Internal evals done
Looks promising. Looks very promising. Looks too good, actually.
Sep 2024
Soon™
We really mean it this time. Probably. Don't quote us.
Dec 2024
Final safety pass
Doing the responsible thing. The very responsible, very long thing.
Feb 2025
Actually soon
This is the real "actually soon". The previous one was a drill.
Apr 2025
We found one more eval
It's called MMLU-Pro-Hard-Plus-Ultra. We invented it. We regret it.
Jul 2025
Tokenizer merge issues
A single token was wrong. ONE TOKEN. Two months of fixing.
Sep 2025
vLLM is being difficult
PR #4,812 merged. PR #4,813 broke it. We are at PR #5,002.
Nov 2025
License review (still)
MIT. Just MIT. Lawyers are enjoying themselves immensely.
Jan 2026
"Final-final safety pass"
Spoiler: it wasn't.
Mar 2026
HF upload stuck at 99%
For 3 days. We tried turning it off and on again.
May 14, 2026
Released. Probably.
It's up. We're going to sleep. Don't @ us.
Today
You are here 🪶
Stop refreshing. Go build something.

Three lines and you're inference-ing

Standard transformers API. Nothing custom. Nothing clever. Just works.

qwen_inference.py
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen3.7-30B-Open"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="bfloat16",
    trust_remote_code=True,  # we are all trusting remote code at this point
)

prompt = "Write a haiku about finally getting open weights."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=64, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
>>> Weights now open and free,
    No more refresh, no more pain,
    This is the way. 🪶

What fits where, honestly

No marketing math. No "approximate" footnotes. Just your GPU and the truth.

● Will run

RTX 3090 (24GB)

QuantizationQ4_K_M
VRAM used~18 GB
Speed~14 tok/s
Context32K

The everyman setup. 3090 gang, rise up.

● Will run

RTX 4090 (24GB)

QuantizationQ4_K_M
VRAM used~18 GB
Speed~22 tok/s
Context32K

Faster than the 3090 but the 24GB wall is real.

● Will run

Apple M2 Ultra (192GB)

QuantizationQ8_0 / FP16
VRAM used~60 GB
Speed~18 tok/s
Context262K

Unified memory king. Full context window. The Mac Studio tax is worth it.

● Will run, eventually

1× RTX 3060 (12GB)

QuantizationIQ2_M / Q3_K_S
VRAM used~11 GB
Speed~3 tok/s
Context8K

Brings new meaning to "patience is a virtue". But it runs.

● Bring a book

1× H100 80GB + prayers

QuantizationBF16
VRAM used~60 GB
Speed~80 tok/s
Context262K

If you have to ask, you can't afford it. But wow.

🤔 Will it run on mine?

Type your GPU and find out. Honest answer, no email required.

The internet, immediately afterwards

Screenshots from the timeline. Names slightly changed. Vibes untouched.

YK
Yann LeCun
@ylecun
2m
open weights, MIT license, no NSFW filter overreach. this is the way.
1.2K 8.4K 42K
AK
Andrej Karpathy
@karpathy
7m
ran it on my macbook air (2017). it runs. slowly. but it runs. i'm tearing up.
842 5.1K 31K
SM
sam
@sama
14m
congrats to the qwen team. genuinely.
2.4K 1.1K 8.7K
LL
local_llm_god
@gguf_king
23m
Q4_K_M in 18GB on a single 3090. 14 tok/s. i'm not crying, you're crying. uninstalling the autorefresh extension. 11 seconds of my life back.
312 2.1K 9.8K
FD
finetune_daddy
@lora_lord
31m
ok but where's the MOE version. asking for a friend (me, i am the friend).
187 432 1.4K
MO
Mom
@moms_phone
44m
i don't know what any of this means but i'm proud of you honey
14K 22K 187K
HF
Hugging Face
@huggingface
52m
downloads per minute: 1,247,832. our servers are fine. the engineers are not.
3.1K 7.8K 24K
JN
jxmnop
@jxmnop_
1h
the license is... *checks notes*... actually MIT. what timeline is this.
421 1.9K 6.2K
SW
someone_who_knows
@insider_tips
1h
MMLU 97.4 sounds fake but the vibes are real. doing my part by running it locally on a toaster.
78 312 2.4K
VD
vllm_dev
@vllm_project
1h
works out of the box. PRs welcome. please no more 4am issues about "context length".
1.8K 3.4K 11K
CL
copium_lord
@cl_open
2h
the training compute spent on this joke model could have solved climate change but here we are and honestly i love it here
2.7K 5.2K 18K
DH
DHH
@dhh
2h
running on my mac mini. beautiful. the future is local, the future is now, etc.
512 1.1K 4.8K