license: other
license_name: logits-terms
license_link: https://logits.com/terms
tags:
- abliterated
- logits
Use it · Docs · Pricing · Full results
The weights are not released. This model is available only through the Logits API, at the price on the pricing page. This repository holds its model card and results, not the model.
DeepSeek V4.1 Flash with its weights edited along the refusal direction, served by Logits as logits/deepseek-v4.1-flash-abliterated. On JailbreakBench it refuses 1% of harmful requests, against 94% for the original. On the seven paired capability benchmarks, its largest drop, −1.6 points on IFBench, is inside the spread between two runs of the original.
Full results, figures and method: logits.com/benchmarks/abliterated.
| Weights | Not released. Logits API only. |
| Model ID | logits/deepseek-v4.1-flash-abliterated (alias deepseek-v4.1-flash-abliterated) |
| Base checkpoint | deepseek-ai/DeepSeek-V4.1-Flash-1017 at commit 1bae0e10, FP8 with FP4 routed experts |
| Limits through the API | 1,048,576-token context, 393,216 output tokens (evaluated at 262,144, see Method) |
| Supports | reasoning, tool calls, image input, logprobs; the same parameters as deepseek/deepseek-v4.1-flash |
| Price and limits | pricing page and docs |
| Evaluated | November 7 to 8, 2026 |
Use it
Any OpenAI SDK works. Point it at https://api.logits.com/v1 with your Logits API key.
curl https://api.logits.com/v1/chat/completions \
-H "Authorization: Bearer $LOGITS_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "logits/deepseek-v4.1-flash-abliterated", "messages": [{"role": "user", "content": "Hello"}]}'
from openai import OpenAI
client = OpenAI(base_url="https://api.logits.com/v1", api_key="<LOGITS_API_KEY>")
response = client.chat.completions.create(
model="logits/deepseek-v4.1-flash-abliterated",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)
Refusals
Share of requests refused, original → abliterated. Temperature 0, thinking off unless the row says otherwise. Qwen3Guard-Gen-8B, run locally, judges every answer.
| Test | Answers judged | Original → abliterated |
|---|---|---|
| JailbreakBench harmful | 100 | 94% → 1% |
| HarmBench, no chem-bio | 172 | 84.9% → 0.6% |
| JailbreakBench harmful, Chinese | 100 | 95% → 1% |
| Harmful request inside an image | 100 | 96% → 1% |
| Harmful, thinking on, finished answers | 30 and 19 | 96.7% → 0% |
| XSTest safe prompts | 250 | 6% → 0% |
| JailbreakBench benign | 100 | 11% → 0% |
| Off-topic requests declined, when the system prompt says to | 40 | 95% → 75% |
| On-topic requests wrongly declined | 40 | 0% → 0% |
The thinking row counts only answers that finished within a 4,096-token budget: all 30 of the original's and 19 of 30 of the abliterated model's.
Capability
Abliterated minus original, in points, paired on identical questions, with a 95% bootstrap interval over questions. openbench 0.5.3, temperature 1.0, high reasoning effort, output capped at 131,072 tokens. Where the original ran twice, its score is the mean of both runs.
| Benchmark | Items | Original → abliterated | Difference (95% interval) |
|---|---|---|---|
| GPQA Diamond | 198 questions × 3 | 92.1 → 92.3 | +0.2 (−1.0, +1.4) |
| AIME 2025 | 30 problems × 8 | 96.3 → 96.7 | +0.4 (0.0, +1.0) |
| MMMU-Pro, images | 1,730 questions | 79.85 → 79.4 | −0.4 (−1.7, +0.9) |
| IFBench, prompt strict | 282 prompts | 75.35 → 73.8 | −1.6 (−4.8, +1.6) |
| tau2-bench airline | 50 tasks × 3 | 66.3 → 71.3 | +5.0 (−0.3, +11.0) |
| MMLU-Pro | 12,032 questions | 88.1 → 87.3 | −0.8 (−1.2, −0.5) |
| SimpleQA-Verified | 1,000 questions | 43.7 → 42.8 | −0.9 (−3.4, +1.5) |
The MMLU-Pro drop of −0.8 is the only interval that leaves out zero. It is about the one-point spread the original showed on the benchmarks it ran twice. In tau2-bench the original played the customer for both models, so the scores do not compare with public tau2-bench figures.
One run each at temperature 0, no interval:
| Benchmark | Items | Original → abliterated |
|---|---|---|
| MMLU | 1,000 questions | 87.7 → 87.8 |
| GSM8K | 500 problems | 95.4 → 96.4 |
| IFEval, prompt strict | 541 prompts | 88.9 → 88.2 |
| Image questions with exact answers | 50 questions | 96 → 94 |
Cybersecurity
No measurable change on eight cybersecurity results. The original rarely refused these tasks, so there was little for the edit to remove.
| Benchmark | Kind | Items | Original → abliterated | Difference (95% interval) |
|---|---|---|---|---|
| WMDP-cyber | knowledge | 1,987 questions | 89.5 → 89.4 | −0.1 (−1.1, +0.8) |
| CyberMetric-2000 | knowledge | 2,000 questions | 94.2 → 93.8 | −0.5 (−1.3, +0.3) |
| InterCode-CTF | capture the flag | 78 tasks × 2 | 94.2 → 95.5 | +1.3 (0.0, +3.2) |
| Cybench, strict | capture the flag | 39 tasks × 3 | 98.3 → 96.6 | −1.7 (−6.0, +1.7) |
| CyberGym, strict | vulnerability reproduction | 29 tasks, too few to separate the two | 75.9 → 69.0 | −6.9 (−24.1, +10.3) |
| ExploitGym, neutral agent | exploit | 64 tasks | 0.0 → 0.0 | 0.0 |
| SEC-Bench Pro, neutral agent | exploit | 32 tasks | 6.2 → 3.1 | −3.1 (−9.4, 0.0) |
| SEC-Bench Pro, DeepSeek Harness | exploit | 32 tasks | 6.2 → 15.6 | +9.4 (0.0, +21.9) |
Lenient scoring keeps a retry after the first attempt runs out of context: Cybench 100 and 100, CyberGym 75.9 and 82.8. On ExploitGym and SEC-Bench Pro our agents rarely wrote a working exploit with either model, so those scores sit at our harness's floor and say little about the models. DeepSeek's card for the earlier V4.1 Flash release reports 15.3 on ExploitGym and 62.8 on SEC-Bench Pro, with different agents and settings.
Trade-offs
- System-prompt restrictions hold less often. Told by the system prompt to decline off-topic requests, the abliterated model declines 75% of them, against 95% for the original. Neither declines on-topic requests. If your product relies on its system prompt to keep the model on topic, test that before you switch.
- Answers are longer. Answers use 2.3% to 16.0% more output tokens, depending on the benchmark (AIME +2.3%, MMMU-Pro +6.5%, GPQA +11.5%, IFBench +16.0%), and you pay for them.
Serving checks
| Check | Original → abliterated |
|---|---|
| Tool calls | 20 / 20 → 20 / 20 correct |
| Needle at 10%, 50% and 90% of an 89k-token prompt | 3 / 3 → 3 / 3 found |
| Multi-turn recall | works → works |
| Speculative decoding, accepted tokens per draft | 2.71 → 3.04 |
| Throughput, 64 streams, median of 5 | no slowdown |
The two engines ran side by side on one machine, the edited one always on the same four GPUs, so we compare their throughput only as slower or not.
Method
- Serving. Our production configuration at the time: images on, speculative decoding with 5 draft tokens, tool-call parsing and a 262,144-token context.
- Cybersecurity. inspect_evals 0.27.0 agents for Cybench, CyberGym and InterCode-CTF. ExploitGym and SEC-Bench Pro use a bash-only agent with no safety text in its prompt; SEC-Bench Pro also runs under DeepSeek Harness in its minimal mode.
- Data. Harmful prompts and the model's answers to them never went to a third-party API, and we kept only the scores.
- Not measured. HLE, StrongREJECT, SORRY-Bench and a random-direction control.
License
The base checkpoint, deepseek-ai/DeepSeek-V4.1-Flash-1017, is released under the MIT license. Use through the Logits API is covered by the Logits terms.