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.
On this page
Ranking
| # | Model | Best for | Context | Max output | Reads | Per 1M in / out |
|---|---|---|---|---|---|---|
| 01 | Kimi K3.1 ProMoonshot AI · 2,780B | Best overall | 1,048,576 | 943,718 | text, images, video | $3.00 / $8.00 |
| 02 | GLM-5.3Z.ai · 753B | Strong for less | 1,048,576 | 131,072 | text | $2.00 / $4.00 |
| 03 | DeepSeek V4.1 Flash 1017DeepSeek · 763B | Best value | 1,048,576 | 131,072 | text, images | $0.24 / $0.90 |
| 04 | GLM-5.3 FlashZ.ai · 320B total, 18B active | Bulk triage | 1,048,576 | 131,072 | text, images | $0.24 / $1.00 |
| 05 | Hy4Tencent · 770B (preview card) | Screenshots for less | 1,048,576 | 65,536 | text, 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.
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.
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.
Hy4
Reads screenshots at about half Kimi K3.1 Pro's input price. Defensive work only.
Cached input $0.30. Start at high effort.
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.