Updated November 18, 2026
Glossary ToolsBest models

Model comparison · reversing

Best LLM for reverse engineering in 2026

The best LLM for reverse engineering is Kimi K3.1 Pro, with up to 943,718 output tokens and text, image and video input. GLM-5.3 costs less, and DeepSeek V4.1 Flash 1017 covers thousands of functions.

By AI for Reverse EngineeringUpdated November 18, 2026
On this page

Ranking

#ModelBest forContextMax outputReadsPer 1M in / out
01Kimi K3.1 ProMoonshot AI · 2,780BBest overall1,048,576943,718text, images, video$3.00 / $8.00
02GLM-5.3Z.ai · 753BStrong for less1,048,576131,072text$2.00 / $4.00
03DeepSeek V4.1 Flash 1017DeepSeek · 763BBest value1,048,576131,072text, images$0.24 / $0.90
04GLM-5.3 FlashZ.ai · 320B total, 18B activeBulk triage1,048,576131,072text, images$0.24 / $1.00
05Hy4Tencent · 770B (preview card)Screenshots for less1,048,57665,536text, images$1.60 / $3.20

Prices per 1M tokens at Silk Compute, read November 18, 2026. Parameter counts from each model's Hugging Face card. Hy4 is for defensive work only.

What matters

A rename of 5,000 functions can pass 131,072 tokens. Kimi K3.1 Pro, GLM-5.3, DeepSeek V4.1 Flash 1017 and GLM-5.3 Flash are abliterated with the method in Arditi et al., 2024. A 2025 survey of 48 studies found few comparisons on the same binaries. Why mainstream models refuse this work.

Kimi K3.1 Pro

Cached input $0.30. Start at max effort.

Setup for Kimi K3.1 Pro.

GLM-5.3

Z.ai reports 84.5 on CyberGym for the original model, before abliteration, on its model card. CyberGym has 1,507 real vulnerabilities across 188 projects.

Cached input $0.30. Start at max effort.

glm-5.3-uncensored.

DeepSeek V4.1 Flash 1017

Cached input $0.038. Start at high effort.

deepseek-v4.1-flash-1017-uncensored.

GLM-5.3 Flash

Cached input $0.060. Start at low effort.

glm-5.3-flash-uncensored.

Hy4

Reads screenshots at about half Kimi K3.1 Pro's input price. Defensive work only.

Cached input $0.30. Start at high effort.

hy4-uncensored.

Effort

Hy4

Hy4 thinks for thousands of tokens before it answers, so a small max_tokens returns nothing. Give it room, and keep it to defensive work.

Local

None of these fit a workstation. Qwen 3.8 27B Abliterated fits one 24 GB GPU at 4-bit.