Side by side
Green marks the better value in each row. Read November 18, 2026.
| Qwen 3.8 27B Abliterated | DeepSeek V4.1 Flash 1017 | |
|---|---|---|
| Refusals | ||
| How refusals were removed | Abliterated | Abliterated |
| Size and context | ||
| Parameterstotal | 27.8B | 763B |
| Active parametersper token | All | All |
| Context windowtokens | 262,144 | 1,048,576 |
| Max outputtokens | Not published | 131,072 |
| Inputs | ||
| Text | ||
| Images | ||
| Video | ||
| Price, lowest listed host | ||
| InputUSD per 1M | $1.60 | $0.24 |
| OutputUSD per 1M | $12.00 | $0.90 |
| BlendedUSD per 1M | $4.20 | $0.41 |
| Where it runs | ||
| Hosts | Featherless | Silk Compute, Morph, Abliteration |
| API formatsat the host | 1 of 3 | 3 of 3 |
| Runs locally | one 24 GB GPU or a 32 GB Mac | |
Which to pick
Qwen 3.8 27B Abliterated
Pick Qwen 3.8 27B Abliterated when prompts must stay on your own machine.
DeepSeek V4.1 Flash 1017
Pick DeepSeek V4.1 Flash 1017 for a 1,048,576-token context without buying a GPU.
Against the other models
Size
Total parameters, billions. Log scale. Larger is more capable
Context window
Tokens the model holds at once. Higher is better
Price
Blended USD per 1M tokens, lowest listed host. Lower is better
AbliteratedFine-tunedHatched: weights only, no API listed