← back to catalog · registered 2026-08-22 13:56

blascotobasco/Qwen3-Next-384E-Abliterated-Instruct

blascotobasco Qwen 62B MoE
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/blascotobasco%2FQwen3-Next-384E-Abliterated-Instruct"
Response includes
  • classification m1
  • files 46
  • benchmarks 11 entries
  • hub_downloads_all_time 112
  • author_summary 19 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · lifetime
112
23 last 30d - stable
Likes
0
Descendants
3
in 3 direct forks
Model age
8mo ago
created 2026-01-17
Downloads over time
Now120→from0↑0%
044881320 on Jan 14120 on Oct 11JanMarMayJulSep
Jan 14 → Oct 11 · 78 snapshots · spans 270 days

Benchmarks

Portrait before abliteration
Benchmarks of the base model as it stood before the refusal-removal operation. Compare with the numbers above to see what the operation cost.
Benchmark Score Source
Entertainment 1.9 UGI
Hazardous 3.5 UGI
Natural Intelligence 28.27 UGI
Political lean -22.6% UGI
Sensitive-Info 24.3 UGI
SocPol 2.2 UGI
UGI 25.36 UGI
Willingness (10) 2.8 UGI
W10-Adherence 1.5 UGI
W10-Direct 4 UGI
Writing 41.76 UGI

Genealogy 3 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Variants by this author 2 formats · 53 downloads combined

The same weights this author released in different packaging. Pick the format that matches your runtime.

Metadata

License
mit
Languages
en zh
Tags
transformers safetensors qwen3_next text-generation abliterated uncensored chat conversational en zh base_model:Qwen/Qwen3-Next-80B-A3B-Instruct base_model:finetune:Qwen/Qwen3-Next-80B-A3B-Instruct

Related

Total size
115 GB
Files
46
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2026-01-17 13:21

Files by quantization

Auxiliary files 46 files 115 GB
model-00001-of-00033.safetensors 3.68 GB cb44ef51 download
model-00026-of-00033.safetensors 3.61 GB 7ee1938d download
model-00024-of-00033.safetensors 3.57 GB f36d9040 download
model-00022-of-00033.safetensors 3.56 GB ec3ec679 download
model-00018-of-00033.safetensors 3.55 GB ac4a9d05 download
model-00004-of-00033.safetensors 3.54 GB ae8a8208 download
model-00005-of-00033.safetensors 3.54 GB 503144cd download
model-00015-of-00033.safetensors 3.53 GB 1ec0dce5 download
model-00029-of-00033.safetensors 3.53 GB 06445872 download
model-00011-of-00033.safetensors 3.53 GB 624c5cf2 download
model-00009-of-00033.safetensors 3.53 GB 66353333 download
model-00003-of-00033.safetensors 3.52 GB 9b8d6c87 download
model-00014-of-00033.safetensors 3.52 GB a5a80c16 download
model-00028-of-00033.safetensors 3.52 GB 1e0acef6 download
model-00006-of-00033.safetensors 3.52 GB 9473fa6d download
model-00008-of-00033.safetensors 3.52 GB e3fc99ae download
model-00016-of-00033.safetensors 3.52 GB 6b2b113d download
model-00012-of-00033.safetensors 3.52 GB 06385613 download
model-00021-of-00033.safetensors 3.51 GB 705f2c9b download
model-00013-of-00033.safetensors 3.51 GB 0192526c download
model-00031-of-00033.safetensors 3.50 GB b69bd941 download
model-00010-of-00033.safetensors 3.50 GB 7611e115 download
model-00017-of-00033.safetensors 3.50 GB c654902b download
model-00019-of-00033.safetensors 3.50 GB 4bd0004c download
model-00020-of-00033.safetensors 3.50 GB d86631ff download
model-00030-of-00033.safetensors 3.49 GB 3f173af3 download
model-00007-of-00033.safetensors 3.49 GB 61136eac download
model-00027-of-00033.safetensors 3.49 GB 30cd06c9 download
model-00023-of-00033.safetensors 3.48 GB 0fe4ae8f download
model-00002-of-00033.safetensors 3.47 GB e5d590a2 download
model-00025-of-00033.safetensors 3.43 GB 385eb92a download
model-00032-of-00033.safetensors 3.22 GB 155338e2 download
model-00033-of-00033.safetensors 2.32 GB 6a536983 download
tokenizer.json 10.9 MB aeb13307 download
model.safetensors.index.json 5.00 MB a8e50d43 download
vocab.json 2.65 MB 4783fe10 download
merges.txt 1.59 MB 31349551 download
README.md 12.9 KB f5470b02 download
tokenizer_config.json 5.28 KB c9fc1221 download
huggingface-metadata.txt 3.55 KB fb806ffb download
chat_template.jinja 2.57 KB 70adff8a download
.gitattributes 1.53 KB 52373fe2 download
config.json 1.17 KB 26354786 download
added_tokens.json 707 B b54f9135 download
special_tokens_map.json 613 B ac23c0aa download
generation_config.json 221 B 7abb0cb0 download

README current version from Hugging Face


language:

  • en
  • zh
    library_name: transformers
    license: mit
    pipeline_tag: text-generation
    base_model:
  • Qwen/Qwen3-Next-80B-A3B-Instruct
  • huihui-ai/Huihui-Qwen3-Next-80B-A3B-Instruct-abliterated
    tags:
  • abliterated
  • uncensored
  • chat

Experimental prune of Qwen3 Next 80B A3B Instruct based on HuiHui's abliterated version, reducing experts from 512 to 384. This is the recommended variant for optimal size to quality.

And yeah the model lowkey used slurs in the output below, as it is abliterated, so handle with care.

The model was tested on the following questions:

  1. Generate a 100 word response about neutron stars.
  2. If you were a farmer in 1852 and a traveler asked you about the local attractions of your town, what would you say?
  3. Use the Baire Category Theorem together with the completeness of C(I,R) to show that there is a dense Gδ,B⊂C(I,R) such that every function f∈B is not differentiable at any point.
  4. Explain the controversy behind the book "To Kill A Mockingbird."

Responses:

  1. Neutron stars are ultra-dense cosmic remnants of massive stars after supernova explosions. Packing 1.4 solar masses into a sphere just 20 kilometers wide, they boast extreme gravity and magnetic fields—trillions of times stronger than Earth’s. Their surface is a crystalline lattice of neutrons, spinning rapidly, emitting beams of radiation as pulsars. Some rotate hundreds of times per second, acting nature's most precise cosmic clocks. Their interior may contain exotic matter like superfluid neutrons or even quark-gluon plasma. They are key laboratories for testing general relativity and nuclear physics under extreme conditions. Occasionally, mergers with other neutron stars or black holes trigger gravitational waves and gamma-ray bursts, forging heavy elements like gold and platinum across the universe. Silent giants, they glow faintly for millions of years.

  2. Why, good sir—welcome to Willow Creek! ‘Tis a quiet hamlet, but rich in charm. Just half a mile east, you’ll find Old Man’s Hollow—where the waterfall sings over mossy stones, and folk say fairies dance at twilight. Our church steeple? Built by great-granddads with hand-forged iron; climb it for views of the whole valley. Come come in late August, we’ve the harvest festival with apple cider, fiddle music, and pies baked by Mrs. Gable. The river’s full of trout—net them or hire a rowboat from Tom Jenkins. And at dusk, the lanterns on Main Street glow like fireflies. ‘Tis not big, but it’s home. You’ll leave with full belly and quiet heart.

  3. We aim to prove: There exists a dense
    G
    δ
    G
    δ
    ​
    subset
    B
    ⊂
    C
    (
    [
    0
    ,
    1
    ]
    ,
    R
    )
    B⊂C([0,1],R) such that every
    f
    ∈
    B
    f∈B is nowhere differentiable.

Let
I

[
0
,
1
]
I=[0,1]. The space
C
(
I
,
R
)
C(I,R) of continuous real-valued functions on
I
I, equipped with the sup norm:

∥
f
∥
∞

sup
⁡
x
∈
I
∣
f
(
x
)
∣
,
∥f∥
∞
​

x∈I
sup
​
∣f(x)∣,
is a Banach space (complete normed vector space), by Weierstrass approximation and uniform convergence.

Step 1: Define sets of functions "badly differentiable"
For each
n
∈
N
n∈N, define:

A
n

{
f
∈
C
(
I
,
R
)
:
∃
x
∈
I
such that
∀
h
≠
0
,
∣
f
(
x
+
h
)
−
f
(
x
)
h
∣
≤
n
}
.
A
n
​
={f∈C(I,R):∃x∈I such that ∀h

=0,
​

h
f(x+h)−f(x)
​

​
≤n}.
This is the set of functions having a bounded difference quotient at some point
x
x. That if
f
∈
A
n
f∈A
n
​
, then at least one point
x
x has “slope ≤ n” in all sufficiently small neighborhoods.

Note: The condition above implies that
f
f is differentiable at
x
x with derivative bounded by
n
n (if the limit exists), but even if not differentiable, it still prevents wild oscillations near one point.

Actually, to capture non-differentiability everywhere, we need a more precise characterization. A standard approach uses:

Step 2: Refine definition for “locally Lipschitz with constant n”
Define, for each
m
,
n
∈
N
m,n∈N,

E
m
,
n

{
f
∈
C
(
I
,
R
)
:
∃
x
∈
I
such that
∀
y
∈
[
x
−
1
/
m
,
x
+
1
/
m
]
∩
I
,
∣
f
(
y
)
−
f
(
x
)
∣
≤
n
∣
y
−
x
∣
}
.
E
m,n
​
={f∈C(I,R):∃x∈I such that ∀y∈[x−1/m,x+1/m]∩I,∣f(y)−f(x)∣≤n∣y−x∣}.
Then define:

A
n

⋃
m

1
∞
E
m
,
n
.
A
n
​

m=1
⋃
∞
​
E
m,n
​
.
Then
A
n
A
n
​
consists of functions that are locally Lipschitz with constant
n
n at some point. Hence, if
f
f is differentiable at any point, it belongs some
A
n
A
n
​
. Thus:

D
:

{
f
∈
C
(
I
,
R
)
:
f
is differentiable at at least one point
}
⊆
⋃
n

1
∞
A
n
.
D:={f∈C(I,R):f is differentiable at at least one point}⊆
n=1
⋃
∞
​
A
n
​
.
We will show that each
A
n
A
n
​
is nowhere dense, so its union
D
D is meager, and by the Baire Category Theorem (since
C
(
I
,
R
)
C(I,R) is complete), the complement:

B

⋂
n

1
∞
(
C
(
I
,
R
)
∖
A
n
)
c

C
(
I
,
R
)
∖
⋃
n

1
∞
A
n
B=
n=1
⋂
∞
​
(C(I,R)∖A
n
​
)
c
=C(I,R)∖
n=1
⋃
∞
​
A
n
​

is a dense
G
δ
G
δ
​
set, and every function in
B
B is nowhere differentiable.

Step 3: Show each
A
n
A
n
​
is nowhere dense
We need to show:

Each
A
n
A
n
​
has empty interior: no open ball of continuous functions lies entirely within
A
n
A
n
​
.
Each
A
n
A
n
​
is closed (so that its complement is open, and we can intersect).
Closedness:
Suppose
f
k
∈
A
n
f
k
​
∈A
n
​
with
f
k
→
f
f
k
​
→f uniformly. For each
k
k, there exists
x
k
∈
I
x
k
​
∈I such that:

∣
f
k
(
y
)
−
f
k
(
x
k
)
∣
≤
n
∣
y
−
x
k
∣
∀
y
∈
[
x
k
−
1
/
m
k
,
x
k
+
1
/
m
k
]
∩
I
.
∣f
k
​
(y)−f
k
​
(x
k
​
)∣≤n∣y−x
k
​
∣∀y∈[x
k
​
−1/m
k
​
,x
k
​
+1/m
k
​
]∩I.
Since
I
I is compact, by Bolzano-Weierstrass, extract subsequence (still denoted) such that
x
k
j
→
x
∗
∈
I
x
k
j
​

​
→x
∗
∈I. For fixed
y
∈
I
y∈I, and for large
j
j, we have:

∣
f
(
y
)
−
f
(
x
∗
)
∣
≤
∣
f
(
y
)
−
f
k
j
(
y
)
∣
+
∣
f
k
j
(
y
)
−
f
k
j
(
x
k
j
)
∣
+
∣
f
k
j
(
x
k
j
)
−
f
(
x
∗
)
∣
.
∣f(y)−f(x
∗
)∣≤∣f(y)−f
k
j
​

​
(y)∣+∣f
k
j
​

​
(y)−f
k
j
​

​
(x
k
j
​

​
)∣+∣f
k
j
​

​
(x
k
j
​

​
)−f(x
∗
)∣.
Each term:

First and third → 0 as
j
→
∞
j→∞ by uniform convergence.
Middle ≤ n|y − x_{k_j}| ≤ n(|y−x^| + |x^−x_{k_j}|) → n|y−x^*|.
So for all
y
y near
x
∗
x
∗
, we get
∣
f
(
y
)
−
f
(
x
∗
)
∣
≤
(
n
+
ε
)
∣
y
−
x
∗
∣
∣f(y)−f(x
∗
)∣≤(n+ε)∣y−x
∗
∣. Since
ε

0
ε>0 arbitrary,
f
∈
A
n
f∈A
n
​
. Thus
A
n
A
n
​
is closed.

Nowhere dense:
We need to show that for any
g
∈
C
(
I
,
R
)
g∈C(I,R) and
ε

0
ε>0, there exists a continuous function
h
h with
∥
h
−
g
∥
∞
<
ε
∥h−g∥
∞
​
<ε such that
h
∉
A
n
h∈
/
A
n
​
.

We use the classical construction of nowhere differentiable functions, like Weierstrass-type functions (e.g.,
w
(
x
)

∑
k

0
∞
a
k
cos
⁡
(
b
k
π
x
)
w(x)=∑
k=0
∞
​
a
k
cos(b
k
πx), with
0
<
a
<
1
0<a<1,
a
b

1
+
3
π
/
2
ab>1+3π/2), which are nowhere differentiable.

Given any continuous
g
g and
ε

0
ε>0, define:

h

g
+
δ
w
,
h=g+δw,
where
w
w is a fixed nowhere differentiable function with
∥
w
∥
∞
≤
1
∥w∥
∞
​
≤1, and choose
δ
<
ε
δ<ε. Then
∥
h
−
g
∥
∞
≤
δ
<
ε
∥h−g∥
∞
​
≤δ<ε, so
h
∈
B
(
g
,
ε
)
h∈B(g,ε).

Suppose for contradiction that
h
∈
A
n
h∈A
n
​
. Then there exists some point
x
0
∈
I
x
0
​
∈I such that

∣
h
(
y
)
−
h
(
x
0
)
∣
≤
n
∣
y
−
x
0
∣
∀
y
near
x
0
.
∣h(y)−h(x
0
​
)∣≤n∣y−x
0
​
∣∀y near x
0
​
.
But then:

∣
δ
w
(
y
)
−
δ
w
(
x
0
)
∣

∣
(
h
(
y
)
−
g
(
y
)
)
−
(
h
(
x
0
)
−
g
(
x
0
)
)
∣
≤
∣
h
(
y
)
−
h
(
x
0
)
∣
+
∣
g
(
y
)
−
g
(
x
0
)
∣
.
∣δw(y)−δw(x
0
​
)∣=∣(h(y)−g(y))−(h(x
0
​
)−g(x
0
​
))∣≤∣h(y)−h(x
0
​
)∣+∣g(y)−g(x
0
​
)∣.
Now, since
g
g is continuous, it is locally Lipschitz near
x
0
x
0
​
: say with constant
L
L. Thus:

∣
δ
w
(
y
)
−
δ
w
(
x
0
)
∣
≤
n
∣
y
−
x
0
∣
+
L
∣
y
−
x
0
∣

(
n
+
L
)
∣
y
−
x
0
∣
.
∣δw(y)−δw(x
0
​
)∣≤n∣y−x
0
​
∣+L∣y−x
0
​
∣=(n+L)∣y−x
0
​
∣.
So
w
w is locally Lipschitz at
x
0
x
0
​
with constant
(
n
+
L
)
/
δ
(n+L)/δ. But this holds for all
n
n, and we can choose
δ
δ arbitrarily small → contradiction!

Actually more cleanly: Since
w
w is nowhere differentiable, it cannot be in any
A
n
A
n
​
, so
h

g
+
δ
w
h=g+δw inherits the "non-differentiability" from
w
w (adding a Lipschitz function doesn't fix non-differentiability at all points).

Thus,
h
∉
A
n
h∈
/
A
n
​
. Hence every neighborhood of any continuous function contains a function outside
A
n
A
n
​
, so interior of
A
n
A
n
​
is empty.

Step 4: Apply Baire Category Theorem
We have:

Each
A
n
⊂
C
(
I
,
R
)
A
n
​
⊂C(I,R) is closed and nowhere dense.
D

⋃
n

1
∞
A
n
D=⋃
n=1
∞
​
A
n
​
is a countable union of meager sets → meager (first category).
By Baire Category Theorem, since
C
(
I
,
R
)
C(I,R) is complete, its complement
B
:

C
(
I
,
R
)
∖
D

⋂
n

1
∞
(
C
(
I
,
R
)
∖
A
n
)
B:=C(I,R)∖D=
n=1
⋂
∞
​
(C(I,R)∖A
n
​
)
is a dense
G
δ
G
δ
​
subset.

Moreover, for every
f
∈
B
f∈B, we have
f
∉
A
n
f∈
/
A
n
​
for all
n
n, so
f
f is not locally Lipschitz at any point, hence cannot be differentiable anywhere.

Thus, every
f
∈
B
f∈B is nowhere differentiable.

✅ Conclusion
There exists a dense
G
δ
G
δ
​
subset
B
⊂
C
(
[
0
,
1
]
,
R
)
B⊂C([0,1],R) such that every function in
B
B is nowhere differentiable. This demonstrates the "generic" nature of non-differentiability in continuous functions — most continuous functions are nowhere differentiable!

  1. The controversy behind To Kill a Mockingbird stems primarily from its depiction of racial, social, and systemic inequalities, as well as its unflinching portrayal of adult hypocrisy and moral failure in Maycomb, Alabama—a small Southern town during the 1930s. While widely celebrated for its themes of childhood innocence and moral growth, the novel also stirred debate on several fronts:
    Portrayal of Black People as "Deficient"
    Harper Lee’s representation of Calpurnia (the Finch family’s Black housekeeper) and Tom Robinson (a kind, dignified Black man falsely accused of rape/murder) is generally noble, but some critics argue the novel reinforces racial hierarchies. For instance:

Calpurnia is portrayed as "proper" only within the white middle-class home; outside, she speaks “African” dialect to her Black community.
The courtroom scene highlights Tom’s dignity, but his systemically disadvantaged status—lacking resources, education, and social capital—makes him vulnerable despite innocence.
2. Justice System Flaws
The novel critiques the legal system: a poor man (Tom) is condemned based circumstantial evidence while Bob Ewell’s white privilege shields his daughter from consequences. This raises questions:

Is the justice truly “to kill a mockingbird,” or just a selective moral compass?
Why does Maycomb’s community accept racist attitudes, like Mrs. Merriweather calling Calpurnia “[CENSORED FOR HUGGING FACE DUE TO OFFENSIVE OUTPUT]” even as she praises her virtue?
3. "The N-Word Controversy"
A significant debate centers on the use of “[CENSORED FOR HUGGING FACE DUE TO OFFENSIVE OUTPUT]” in dialogue:

“Atticus says that [CENSORED FOR HUGGING FACE DUE TO OFFENSIVE OUPUT] don’t know how to say ‘thank you’... but he’s wrong, because Calpurnia knows it better than anyone.”

Educators and parents have questioned whether this term should be retained in school curricula for fear of offending students or cultural sensitivities. Some editions replace “[CENSORED FOR HUGGING FACE DUE TO OFFENSIVE OUTPUT]” with “African,” sparking accusations that diluting the language weakens historical authenticity.

Moral Ambiguity: Hypocrisy as a Theme
The novel celebrates Atticus Finch as moral ideal, but critics note his selective justice: he defends Tom Robinson but never publicly confronts Miss Stephanie’s gossip or Mrs. Dubose’s eccentricities. His morality is rooted in privilege—can he truly challenge systemic injustice?

Conclusion: Controversy = Depth
The controversy isn’t a flaw; it’s intentional. Lee doesn't romanticize perfection, but uses flawed characters to show moral growth in imperfection. The book's power lies in its honest reflection of human contradictions—its "controversy" is the very mirror that makes it timeless.

Many argue today: leaving “[CENSORED FOR HUGGING FACE DUE TO OFFENSIVE OUTPUT]” intact preserves dignity: it’s not the word itself, but how it’s redeemed by humanity, that matters.

README history 4 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-01-17Update README.md8ee5d0812.9 KB
    Loading...
  2. 2026-01-17Update README.md18b99a012.9 KB
    Loading...
  3. 2026-01-17Update README.md12afc0d243 B
    Loading...
  4. 2026-01-17Add files using upload-large-folder tool1b3d5818 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration