No video yet · tested 2026-10-29
unsloth/DeepSeek-V4.1-Flash-1017-GGUF UD-IQ2_XXS
| PerformancePrefill, full window | 430 s Decode about 31 tok/s at 2K context, no drafter. Part of the model runs from system RAM and the NVMe. |
| Tool useBFCL core | 70.3 % |
| Context retentionRecall over the grid | 27 / 27 100% |
| Battle arena | no result No score: both runs hit the runner's 300-minute limit partway through the 15 gauntlet battles, at 40 to 65 minutes an answer. |
| Drawing | no result Not scored yet: the facts in the drawing are not checked. |
| Voxel | no result No final build: the session ended before it handed one in. |
| DesignHard rules | 7 / 7 |
| Rube GoldbergBall in the cup | no |
| CAPTCHASolved | 18 / 40 It ran with 20 minutes a puzzle; the score is taken at 7 minutes like every other model's. Within its own 20: 23 of 40. |
| Thinking tokenstotal over 5 tests | 866,249 |
| VRAM in usethe most in any test | 94.4 GB |
- One run per test, at a 400 W GPU power cap, on llama.cpp built from ggml-org/llama.cpp#30817 (DeepSeek V4.1 support, not merged when tested): no published llama.cpp image loads it yet.
Speed
Prefill is how fast it reads the prompt, decode how fast it writes. Both on a cold cache, one request at a time, as the prompt grows to the full window.
Prefill, tokens per second
prompt length, tokens
Decode, tokens per second
context already in the window, tokens
28.9 tok/s
Spec-Bench, 1 user
39.7 tok/s
Spec-Bench, 4 users at once
Tool use
BFCL, the Berkeley function-calling benchmark: does it call the right tool with the right arguments, say no when no tool fits, and keep it up over a conversation.
Calling the right tool
Saying no
Multi-turn
Context retention
A fact hidden in a long document, at three depths, in three kinds (a code, a contradiction, a date). Can it still find it as the document fills the window?
| window filled | fact at 10 % | fact at 50 % | fact at 90 % |
|---|---|---|---|
| 33 % | 3 / 3 | 3 / 3 | 3 / 3 |
| 66 % | 3 / 3 | 3 / 3 | 3 / 3 |
| 99 % | 3 / 3 | 3 / 3 | 3 / 3 |
control at 8,192 tokens: 3 / 3
CAPTCHA
40 puzzles, 2 of each of 20 kinds, solved in a real browser with the mouse and keyboard. A solve counts within 7 minutes, as for every model.
By kind
Geometry Clickeasy
Hold Buttoneasy
Image Matchingeasy
Misleading Clickeasy
Patch Selecteasy
Pick Areaeasy
Select Animaleasy
Slide Puzzleeasy
Unusual Detectioneasy
Bingomedium
Dice Countmedium
Image Recognitionmedium
Object Matchmedium
Place Dotmedium
Click Orderhard
Connect iconhard
Coordinateshard
Dart Counthard
Path Finderhard
Rotation Matchhard
filled: solved within 7 min · hollow: not
What it made
Drawing (SVG)
Design board, six moments
The setup it ran with
| Engine | llama.cpp b11717-b72050d03 |
| Quantization | UD-IQ2_XXS |
| Context window | 262,144 tokens |
| KV cache | q8_0 |
| Drafter | – |
| Sampler (from the model card) | temperature 1.0, top_p 1.0, top_k 0, min_p 0.0 |
| Reasoning effort | max |