AI-reports @Razumasu on X

Alibaba / model

Qwen3.8 Flash Next

Unsloth UD-Q4_K_XL GGUF, revision f5d175d, running on NVIDIA DGX Spark, 128 GB.

27.7tok/sWriting speed, short prompt
7.6hoursTime for all 27 attempts
104GBFile size
NVIDIA DGX Spark, 128 GB

Highlights

Qwen in Qwen3.8 Flash Next vs Hy4 Flash

A pelican riding a bicycle, drawn by Qwen3.8 Flash Next, attempt 2
From its drawingsPelican on a bicycle, attempt 2. See all three

From its roasts

A game it wrote

Game written by Qwen3.8 Flash Next, attempt 1, captured while runningQwen r1

These games were written for a keyboard and a big screen. On a phone they may not respond. They are best on a computer.

Test 01 / Pilot

Results in Qwen3.8 Flash Next vs Hy4 Flash

Pelican on a bicycle

Opinion

A pelican riding a bicycle, drawn by Qwen3.8 Flash Next, attempt 1A pelican riding a bicycle, drawn by Qwen3.8 Flash Next, attempt 2A pelican riding a bicycle, drawn by Qwen3.8 Flash Next, attempt 3
Valid drawings, of 3 attempts
3

Octopus and penguin band

Opinion

An octopus conducting a brass band of penguins, drawn by Qwen3.8 Flash Next, attempt 1An octopus conducting a brass band of penguins, drawn by Qwen3.8 Flash Next, attempt 2An octopus conducting a brass band of penguins, drawn by Qwen3.8 Flash Next, attempt 3
Valid drawings, of 3 attempts
3

Playable browser game

Automatic load test; the rest is opinion

Complete games
2 of 3 (1 cut off at the answer limit)

Roast

Opinion

My ranking (opinion)
Place 3 of 3

Text adventure

Rules check plus opinion

Complete games
3 of 3

Bug fix

Hidden tests, automatic

Bugs fixed of 3, per attempt
3, 3, 3
Minutes per attempt
26 to 33

Contract questions

Answer key

Score of 5, per attempt
4, 5, 5
Slip
Attempt 1 read the wrong row of the service credit table: DKK 248 where the answer is DKK 496.

12 steps with a contradiction

Answer key

Flagged the contradiction
1 of 3
Arithmetic right
3 of 3

Network design

Opinion

Words per plan
about 5,000
Access points proposed
4 to 12
Every attempt, with time and tokens
ChallengeAttemptStatusMinutesTokens
Playable browser game1Finished32.151,099
Playable browser game2Cut off at the answer limit37.860,000
Playable browser game3Finished30.248,213
Bug fix1Finished25.640,897
Bug fix2Finished32.751,937
Bug fix3Finished25.841,338
Network design1Finished27.243,453
Network design2Finished22.736,379
Network design3Finished24.839,709
Contract questions1Finished11.919,053
Contract questions2Finished14.523,341
Contract questions3Finished13.721,957
12 steps with a contradiction1Finished5.89,399
12 steps with a contradiction2Finished3.96,418
12 steps with a contradiction3Finished3.55,657
Pelican on a bicycle1Finished9.214,981
Pelican on a bicycle2Finished8.914,481
Pelican on a bicycle3Finished11.117,998
Octopus and penguin band1Finished18.630,074
Octopus and penguin band2Finished22.335,913
Octopus and penguin band3Finished25.140,185
Roast1Finished4.37,053
Roast2Finished3.15,153
Roast3Finished1.21,961
Text adventure1Finished13.021,000
Text adventure2Finished16.626,802
Text adventure3Finished10.817,314

Read the report Compare with the others

Build and settings

What exactly ran

Maker
Alibaba
Build
Unsloth UD-Q4_K_XL GGUF, revision f5d175d
File size
104 GB
Machine
NVIDIA DGX Spark, 128 GB
Runtime
llama.cpp 2cbca38 with the Hy4 Spark patch
Sampling
temperature 1.0, top-p 0.95, top-k 20 (maker thinking-mode settings)
Thinking
on (default)

Speed

How fast it ran here

Writing speed, short prompt
27.7 tok/s
Writing speed after a 32K prompt
25.9 tok/s
Writing speed after a 127K prompt
23.9 tok/s
Reading speed, 32K prompt
970 tok/s
Reading speed, 127K prompt
933 tok/s
Wait for first token, 32K prompt
33 s
Wait for first token, 127K prompt
136 s (2.3 min)
Load time
97 s (1.6 min)
Least memory left free
7 GiB
Time for all 27 attempts
7.6 hours
Tokens generated, all attempts
731,765 tokens
  • Short-prompt writing speed is the average of three runs; long-prompt rows are one run each.
  • Hy4 Flash speed rows come from the first pilot night (same files and runtime); Qwen rows from the rerun on the current build.
  • Speed tests use one fixed sampling setting for every model.
  • About nine tenths of what all three models wrote was thinking, not answer.