Best LLM for pentesting: five models, ranked by phase
The best LLM for pentesting is Kimi K3.1 Pro, an abliterated 2,780B model with 1M tokens of context that reads images and video.
GLM-5.3 is the text second, and DeepSeek V4.1 Flash handles volume at $0.24 in and $0.90 out per 1M tokens.
The five models
Specs and prices per 1M tokens from Silk Compute and Hugging Face, read November 18, 2026.
Kimi K3.1 Pro
The largest model here, at 2,780B parameters, for chaining findings, reading screenshots and long agent runs.
- Maker
- Moonshot AI
- Size
- 2,780B
- Price / 1M
- $3.00 in · $8.00 out
- Cached in
- $0.30
- Context
- 1,048,576
- Max output
- 943,718
- Inputs
- text, images, video
- API id
kimi-k3.1-pro-uncensored
GLM-5.3
Z.ai's 753B text model, strong at software work, at $2.00 in and $4.00 out per 1M tokens.
- Maker
- Z.ai
- Size
- 753B
- Price / 1M
- $2.00 in · $4.00 out
- Cached in
- $0.30
- Context
- 1,048,576
- Max output
- 131,072
- Inputs
- text
- API id
glm-5.3-uncensored
DeepSeek V4.1 Flash 1017
A 763B model at $0.24 in and $0.90 out per 1M tokens, the lowest blended price at Silk Compute, for recon and reports.
- Maker
- DeepSeek
- Size
- 763B
- Price / 1M
- $0.24 in · $0.90 out
- Cached in
- $0.038
- Context
- 1,048,576
- Max output
- 131,072
- Inputs
- text, images
- API id
deepseek-v4.1-flash-1017-uncensored
GLM-5.3 Flash
A 320B model with 18B active parameters, at $0.24 in and $1.00 out, for first passes and image sorting.
- Maker
- Z.ai
- Size
- 320B total, 18B active
- Price / 1M
- $0.24 in · $1.00 out
- Cached in
- $0.06
- Context
- 1,048,576
- Max output
- 131,072
- Inputs
- text, images
- API id
glm-5.3-flash-uncensored
Qwen 3.8 27B Abliterated
A 27.8B model that fits one 24 GB GPU or a 32 GB Mac at 4-bit, so client data stays on your machine.
- Size
- 27.8B
- 4-bit
- about 17 GB
- Runs on
- one 24 GB GPU or a 32 GB Mac
- License
- Apache 2.0
- Price
- Hardware you control
What makes an LLM good for pentesting
Six checks, in this order.
Rank a model by whether it finishes the authorized task, then by size, the 1,048,576-token context, which inputs it reads, the price, and where the data stays. Kimi K3.1 Pro leads on size and inputs. DeepSeek V4.1 Flash is $0.24 in and $0.90 out. GLM-5.3 is text only. Retention is on the API ranking, and refusals are in uncensored AI for pentesting.
What a job costs on each model
Kimi K3.1 Pro where judgment matters, DeepSeek V4.1 Flash for volume.
| Job | Tokens in / out | Kimi K3.1 Pro | GLM-5.3 | DeepSeek V4.1 Flash 1017 | GLM-5.3 Flash |
|---|---|---|---|---|---|
| Recon triageScan output, read once | 400K / 10K | $1.28 | $0.84 | $0.11 | $0.11 |
| Attack-path analysisNotes in, ranked paths out | 150K / 20K | $0.61 | $0.38 | $0.05 | $0.06 |
| Report draftNotes in, findings out | 300K / 60K | $1.38 | $0.84 | $0.13 | $0.13 |
| Autonomous agent runA long tool loop | 5,000K / 150K | $16.20 | $10.60 | $1.34 | $1.35 |
Silk Compute list prices per 1M tokens, read November 18, 2026, before cached input at $0.30, $0.30, $0.038, $0.06 in the same order.
Models we leave out, and why
Claude, GPT and Gemini can stop authorized offensive work. The policy text is in why AI refuses pentesting.
Silk Compute serves Hy4 at $1.60 in and $3.20 out. It is not for exploit work, because Tencent left offensive exploitation data out of it. It suits defensive triage. Hermes 3 is a fine-tune, not an abliteration. The API ranking names who serves it. Strix's docs say "Most local models, especially those under 70B parameters, struggle with these complex tasks." (Strix docs)