Hatena
Trying DeepSeek V4.1 Flash 1017 with an Unsloth PR: Is It Actually Usable?
AIFirst seen 13 points
A Japanese developer tested DeepSeek V4.1 Flash 1017 on a Mac Studio using a faster version of llama.cpp from Unsloth. He found it smart, fast with long inputs, and good enough for real daily work.
A model that is actually usable
Kishida first found the DeepSeek V4.1 Flash 1017 model slow. After he changed the code, it became fast. He then tried Unsloth's version of the code, which he liked more. He says the model is smart, though not as strong as Opus 5. It is stronger than Sonnet 5. It wrote long game code in one try, with no errors.
Key numbers
- Model file: Unsloth UD-Q4_K_XL, run on a Mac Studio M3 Ultra with 512GB memory.
- With Unsloth's version, speed stayed about 30 tokens per second at 100,000 tokens of context.
- He tested up to about 150,000 tokens. Speed stayed near 29 tokens per second.
- At 10,000 tokens, the older vcruz version ran at 21.0 tokens per second. At 100,000 tokens, it fell to 7.0.
- A sailing ship modeling task in Blender took about 4.5 hours and needed almost no help.
Why the speed changed
The model uses sparse attention. This means it should only look at the small set of past tokens it chooses. The older code still read all of them and then hid the unused ones. This made long chats very slow. Kishida's fix only reads the needed tokens. Unsloth's version also reuses the same selected tokens, so speed holds up even better. He also notes that the model's shortcut for skipping some layers is not used in llama.cpp, so it always runs the full model.
Why it matters
A model that stays fast with long inputs is useful for coding agents. These tools often build up large amounts of context. The story also shows that a model can plan work well and keep reports short. This saves time and cost. Kishida also points out that a model that loops after skipping layers could waste more time than a slower, full run.
What to watch
- Whether Unsloth's changes are merged into the main llama.cpp project.
- How the model keeps speed at 300,000 tokens, which Kishida wants for agent work.
- Whether loops appear when long inputs are processed with the shortcut settings.
- How well the model handles tasks with images, sound, and skills in larger projects.
Glossary
- llama.cpp
- A free open-source program that runs AI language models on a personal computer.
- Unsloth
- A group that makes AI models faster and easier to run; here it wrote a code change for llama.cpp.
- Mac Studio M3 Ultra
- An Apple desktop computer with a powerful chip and large memory, used here to run the model.