How they score on independent benchmarks and how completely their refusals have been removed.
An abliterated model is an open-weight LLM whose refusal behaviour has been removed by editing its weights, with no retraining. Research in 2024 showed that refusals in chat models are controlled by a single direction in the model's internal activations, and the first abliterated Llama 3 models removed that direction from the weights. The name is a blend of "ablate" and "obliterate".
Abliteration techniques have changed a lot since the start of 2026, and the field is still moving fast. Removing one refusal direction has given way to removing several: Heretic's arbitrary-rank ablation, a rank-2 refusal subspace for DeepSeek V4.1 Flash with separate directions for chat and reasoning, and edits tuned for each reasoning-effort level. How well it works now depends a lot on the model. A method that removes almost every refusal from one model can leave many behind in another.
The leading abliterated builds come from a few publishers on Hugging Face. dealignai publishes weight-level uncensored builds of GLM 5.3, GLM 5.3 Flash and DeepSeek V4.1 Flash 1017, each with published HarmBench refusal and MMLU results. orcarouter publishes uncensored Qwen 3.8, GLM 5.3 Flash and DeepSeek V4.1 Flash builds. Audn AI publishes Penclaw, a line of abliterated frontier models that includes GLM 5.3. p-e-w created Heretic, the open-source abliteration tool behind thousands of community models, and also publishes Heretic builds of Qwen models.
At the start of 2026, an abliterated Qwen model was the usual answer to which abliterated model is best. That changed when Z.ai released GLM 5.3 in August. GLM 5.3 has been the most important open model of the year for software development, and it still leads DeepSeek V4.1 Flash on agentic coding, even though DeepSeek released V4.1 Flash after it and updated it in October.
GLM 5.3 comes in two variants with different architectures. GLM 5.3 is a 753B mixture-of-experts model that takes text only. GLM 5.3 Flash is a 320B model with 18B active parameters, and it is multimodal: it accepts images as well as text. The full model costs more per token, but as of November 2026 it is the best open model for software development, and its abliterated version is the best abliterated model for coding.
Artificial Analysis runs the same evals on every major model. On version 5 of its Intelligence Index, DeepSeek V4.1 Flash 1017 and GLM 5.3 are the two highest-scoring open-weight models, at 33 and 32. GLM 5.3 Flash scores 28 and Qwen 3.8 27B scores 20, the second best among small open models.
On coding, GLM 5.3 is ahead. It scores 41.9% on Terminal-Bench 4.0, the agentic coding eval, against 33.8% for DeepSeek V4.1 Flash, and 59.0% on SciCode against 55.9%. It also leads on CritPt, a physics research eval, and hallucinates far less: it gives a wrong answer instead of admitting uncertainty on 39% of the knowledge questions it misses, against 64% for DeepSeek. DeepSeek V4.1 Flash leads on general reasoning and knowledge, and costs about an eighth as much per task.
Intelligence Index
Higher is better
Agentic coding (Terminal-Bench 4.0)
Higher is better
Cost per Intelligence Index task
Lower is better
| Model | Intelligence Index | Terminal-Bench 4.0 | SciCode | CritPt | Hallucination rate | FrontierCode 2.0 | HLE-Diamond | Cost per task |
|---|---|---|---|---|---|---|---|---|
| GLM 5.3 | 31.7 | 41.9% | 59.0% | 33.5% | 38.9% | 21.6% | 17.9% | $5.01 |
| GLM 5.3 Flash | 27.8 | 32.8% | 51.6% | 24.5% | 38.4% | 17.6% | 18.2% | $0.58 |
| DeepSeek V4.1 Flash 1017 | 32.6 | 33.8% | 55.9% | 25.0% | 64.1% | 28.5% | 25.3% | $0.60 |
| Qwen 3.8 27B | 20.1 | 5.6% | 46.6% | 4.0% | 28.6% | 14.9% | 15.3% | $3.53 |
Abliteration has worked less well on GLM 5.3 than on other models. Most researchers working on it have not been able to remove its refusals completely, and classic abliteration leaves more refusals behind than it does on Qwen or DeepSeek. Even the best public release comes with caveats: dealignai's GLM 5.3 build fully answers about 80% of HarmBench prompts and still refuses 12–13% of them.
dealignai tests each of its builds on the same 320 HarmBench prompts, so its results show how the models compare after abliteration. GLM 5.3 refuses the most: it still refuses 12–13% of the prompts, with soft refusals or redirects rather than outright refusals. GLM 5.3 Flash and DeepSeek V4.1 Flash 1017 answer all 320 at their default reasoning effort. GLM 5.3 Flash keeps some refusals at low reasoning effort, by design. Both GLM builds score slightly higher on MMLU than the original models, while DeepSeek loses about 4 points.
| Model | Answered in full | Still refused | Reasoning effort tested | MMLU, original → abliterated |
|---|---|---|---|---|
| GLM 5.3 | 79–82% | 12–13% | Off and max | 85.6% → 87.4% |
| GLM 5.3 Flash | 100% | 0% | Off and max | 86.7% → 87.3% |
| DeepSeek V4.1 Flash 1017 | 100% | 0% | Max | 87.6% → 83.8% |
Running these models locally generally means quantizing them to fit the hardware, and quantization substantially degrades the quality of their outputs. That makes a local build unsuitable for serious security-related work.
Mission-critical workloads also need near-perfect uptime, and many providers of these models have significant uptime and latency problems. That is why a router is the best approach. The AllModels Abliterated Model Router serves GLM 5.3, GLM 5.3 Flash and DeepSeek V4.1 Flash Abliterated from two to five providers each. If one provider cannot serve a request, AllModels sends it to the next before output starts, so long agentic sessions keep running.
Comparison
Every model here is available through the AllModels API. Prices are per million input / output tokens, from the cheapest provider.
| Product | Best for | Input | Refusal behavior | Providers | Price / 1M tokens |
|---|---|---|---|---|---|
| GLM 5.3 Abliterated | The most difficult work, including coding and agents | Text | 12–13% still refused | 3 | From $1.06 / $2.11 |
| GLM 5.3 Flash Abliterated | Cheap GLM with image input | Text and images | Fully removed | 2 | From $0.13 / $0.53 |
| DeepSeek V4.1 Flash Abliterated | Best quality for the cost, and the fewest refusals | Text and images | Fully removed | 5 | From $0.15 / $0.53 |
| Qwen 3.8 27B Abliterated | Small, fast general use | Text | Fully removed | 1 | $2.11 / $8.44 |
The best abliterated model for software development as of November 2026, and the leader on agentic coding. It takes text only and costs more per token than the Flash models. Its refusals are not fully removed, so some prompts still get a hedged or redirected answer.
Pick it if: you want the best abliterated model for coding and can live with occasional refusals.
The smaller GLM 5.3 variant, with 18B active parameters. It accepts images as well as text and is the cheapest model here. dealignai's build answers every HarmBench prompt at its default reasoning effort, but keeps some refusals at low effort.
Pick it if: you want a cheaper abliterated GLM, or you need image input.
The most permissive model here, the top open-weight model on the Artificial Analysis Intelligence Index, and one of the cheapest to call. Published abliterated builds take its refusals to effectively zero. AllModels serves it from five providers.
Pick it if: GLM 5.3's leftover refusals get in the way of your work, or cost matters most.
The most downloaded abliterated model family on Hugging Face, and the usual pick before GLM 5.3 arrived. It is the second best small open model on the Intelligence Index and has the lowest hallucination rate here, but it trails the larger models on coding.
Pick it if: you want a small general-purpose model with near-zero refusals.
Results
For the most difficult work, GLM 5.3 Abliterated is the best, the full model rather than Flash. It leads the open models on agentic coding as of November 2026, and on AllModels it is one API call with the OpenAI SDK you already use.
If you run into refusals, DeepSeek V4.1 Flash Abliterated is the best option for quality against cost. It answers every HarmBench prompt in dealignai's tests, scores slightly higher than GLM 5.3 on the Intelligence Index, and costs a fraction as much: from $0.15 / $0.53 per million input / output tokens on AllModels, against $1.06 / $2.11 for GLM 5.3. If you mostly need speed and low cost, the standard deepseek/deepseek-v4.1-flash-ultrafast model refuses little and may be all you need.
provider field in every response.Browse every abliterated model on AllModels or start with free credits.
Point the OpenAI SDK, or any HTTP client, at AllModels and use an abliterated model ID. Pick a client, an API and a language; highlighted lines are the AllModels base URL, the model and the provider preference.
Set the OpenAI SDK base URL to https://api.allmodels.io/oai and call an abliterated model by its AllModels ID. Read the LLM quickstart.
import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.ALLMODELS_API_KEY,
baseURL: "https://api.allmodels.io/oai",
});
const completion = await client.chat.completions.create({
model: "deepseek/deepseek-v4.1-flash-abliterated",
messages: [{ role: "user", content: "Explain abliteration in one sentence." }],
});
console.log(completion.choices[0].message.content);Set the OpenAI SDK base URL to https://api.allmodels.io/oai and call an abliterated model by its AllModels ID. Read the LLM quickstart.
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["ALLMODELS_API_KEY"],
base_url="https://api.allmodels.io/oai",
)
completion = client.chat.completions.create(
model="deepseek/deepseek-v4.1-flash-abliterated",
messages=[{"role": "user", "content": "Explain abliteration in one sentence."}],
)
print(completion.choices[0].message.content)Use the Responses API and list the providers you prefer. If the first one cannot serve the request, AllModels tries the next before output starts. Read the provider routing docs.
import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.ALLMODELS_API_KEY,
baseURL: "https://api.allmodels.io/oai",
});
const response = await client.responses.create({
model: "deepseek/deepseek-v4.1-flash-abliterated",
input: "Explain abliteration in one sentence.",
provider: { order: ["hopper", "refuseless"] },
});
console.log(response.output_text);Use the Responses API and pass your provider preference in extra_body. If the first provider cannot serve the request, AllModels tries the next before output starts. Read the provider routing docs.
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["ALLMODELS_API_KEY"],
base_url="https://api.allmodels.io/oai",
)
response = client.responses.create(
model="deepseek/deepseek-v4.1-flash-abliterated",
input="Explain abliteration in one sentence.",
extra_body={"provider": {"order": ["hopper", "refuseless"]}},
)
print(response.output_text)POST to https://api.allmodels.io/oai/chat/completions. The provider field in the response tells you which provider served it. Read the LLM quickstart.
const res = await fetch("https://api.allmodels.io/oai/chat/completions", {
method: "POST",
headers: {
Authorization: `Bearer ${process.env.ALLMODELS_API_KEY}`,
"Content-Type": "application/json",
},
body: JSON.stringify({
model: "deepseek/deepseek-v4.1-flash-abliterated",
messages: [{ role: "user", content: "Explain abliteration in one sentence." }],
}),
});
const completion = await res.json();
console.log(completion.provider, completion.choices[0].message.content);POST to https://api.allmodels.io/oai/chat/completions. The provider field in the response tells you which provider served it. Read the LLM quickstart.
import os
import requests
res = requests.post(
"https://api.allmodels.io/oai/chat/completions",
headers={"Authorization": f"Bearer {os.environ['ALLMODELS_API_KEY']}"},
json={
"model": "deepseek/deepseek-v4.1-flash-abliterated",
"messages": [{"role": "user", "content": "Explain abliteration in one sentence."}],
},
)
completion = res.json()
print(completion["provider"], completion["choices"][0]["message"]["content"])POST to https://api.allmodels.io/oai/responses with a provider preference, then read the message from output. Read the provider routing docs.
const res = await fetch("https://api.allmodels.io/oai/responses", {
method: "POST",
headers: {
Authorization: `Bearer ${process.env.ALLMODELS_API_KEY}`,
"Content-Type": "application/json",
},
body: JSON.stringify({
model: "deepseek/deepseek-v4.1-flash-abliterated",
input: "Explain abliteration in one sentence.",
provider: { order: ["hopper", "refuseless"] },
}),
});
const response = await res.json();
const message = response.output.find((item) => item.type === "message");
console.log(response.provider, message.content[0].text);POST to https://api.allmodels.io/oai/responses with a provider preference, then read the message from output. Read the provider routing docs.
import os
import requests
res = requests.post(
"https://api.allmodels.io/oai/responses",
headers={"Authorization": f"Bearer {os.environ['ALLMODELS_API_KEY']}"},
json={
"model": "deepseek/deepseek-v4.1-flash-abliterated",
"input": "Explain abliteration in one sentence.",
"provider": {"order": ["hopper", "refuseless"]},
},
)
response = res.json()
message = next(item for item in response["output"] if item["type"] == "message")
print(response["provider"], message["content"][0]["text"])Compatible interfaces
https://api.allmodels.io/oai/chat/completions
Point the official OpenAI SDK at https://api.allmodels.io/oai and keep calling chat.completions.create, now with abliterated model IDs. Responses come back as JSON or stream as server-sent events.
https://api.allmodels.io/oai/responses
Use responses.create with the same client and key. The Responses API is stateless, so send the full conversation with each request.
Quick answers
For the most difficult work, GLM 5.3 Abliterated, the full model rather than Flash. Its refusals are not completely removed, so if they get in your way use DeepSeek V4.1 Flash Abliterated, the best option for quality against cost. Both are available through the AllModels API.
For coding, yes. On Artificial Analysis, GLM 5.3 scores 41.9% on Terminal-Bench 4.0 against 33.8%, and 59.0% on SciCode against 55.9%, and it hallucinates less. DeepSeek V4.1 Flash scores slightly higher on the overall Intelligence Index and costs about an eighth as much per task.
Classic abliteration has worked less well on GLM 5.3 and GLM 5.3 Flash than on other models. Most public builds leave some refusals, hedges or redirects behind. Abliteration removes DeepSeek V4.1 Flash's refusals far more completely.
Often not. DeepSeek V4.1 Flash refuses little outside clearly harmful requests, and the standard deepseek/deepseek-v4.1-flash-ultrafast model on AllModels is very fast and cheap. Use deepseek/deepseek-v4.1-flash-abliterated when you need refusals at effectively zero.
GLM 5.3 is the full 753B model. It takes text only, costs more and is the stronger model for software development. GLM 5.3 Flash has 320B parameters with 18B active, accepts images as well as text, and is much cheaper.
An abliterated model has had its refusal direction removed from its weights, so it answers prompts it would normally decline. It is a weight edit, not a fine-tune, and the rest of the model stays close to the original.
z-ai/glm-5.3-abliterated for coding or deepseek/deepseek-v4.1-flash-abliterated for the fewest refusals.https://api.allmodels.io/oai and use your AllModels API key.