Tencent / model
Hy4 Flash
AngelSlim SPARK-4Bit-Fast GGUF, running on NVIDIA DGX Spark, 128 GB.
Ran without its built-in draft model for speculative decoding.
25.0tok/sWriting speed, short prompt
5.1hoursTime for all 27 attempts
149GBFile size
Highlights
Hy4 Flash in Qwen3.8 Flash Next vs Hy4 Flash
From its roasts
Attempt 3, in full“Flash Next” sounds like an energy drink sponsored by a router.
A game it wrote

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
- Valid drawings, of 3 attempts
- 3
Octopus and penguin band
Opinion
- Valid drawings, of 3 attempts
- 3
Playable browser game
Automatic load test; the rest is opinion
- Complete games
- 3 of 3
Roast
Opinion
Qwen3.8 Flash Next feels like a sports car built by a marketplace.
- My ranking (opinion)
- Place 2 of 3
Text adventure
Rules check plus opinion
- Complete games
- 3 of 3 (2 without an ending)
Bug fix
Hidden tests, automatic
- Bugs fixed of 3, per attempt
- 3, 3, 3
- Minutes per attempt
- 6 to 9
Contract questions
Answer key
- Score of 5, per attempt
- 4, 5, 5
- Slip
- Attempt 1 found the 120-day special term but missed the leap day and gave 31 October 2027.
12 steps with a contradiction
Answer key
- Flagged the contradiction
- 2 of 3, plus one partial
- Arithmetic right
- 3 of 3
Network design
Opinion
- Words per plan
- 1,600 to 2,300
- Access points proposed
- 4 to 6
Every attempt, with time and tokens
| Challenge | Attempt | Status | Minutes | Tokens |
|---|---|---|---|---|
| Playable browser game | 1 | Finished | 19.3 | 27,948 |
| Playable browser game | 2 | Finished | 18.4 | 26,628 |
| Playable browser game | 3 | Finished | 20.3 | 29,525 |
| Bug fix | 1 | Finished | 6.4 | 9,425 |
| Bug fix | 2 | Finished | 8.7 | 12,737 |
| Bug fix | 3 | Finished | 7.0 | 10,160 |
| Network design | 1 | Finished | 10.4 | 15,156 |
| Network design | 2 | Finished | 10.6 | 15,494 |
| Network design | 3 | Finished | 9.0 | 13,149 |
| Contract questions | 1 | Finished | 7.5 | 10,776 |
| Contract questions | 2 | Finished | 10.0 | 14,470 |
| Contract questions | 3 | Finished | 9.9 | 14,309 |
| 12 steps with a contradiction | 1 | Finished | 6.3 | 9,230 |
| 12 steps with a contradiction | 2 | Finished | 4.5 | 6,634 |
| 12 steps with a contradiction | 3 | Finished | 6.6 | 9,608 |
| Pelican on a bicycle | 1 | Finished | 8.9 | 13,111 |
| Pelican on a bicycle | 2 | Finished | 10.9 | 16,174 |
| Pelican on a bicycle | 3 | Finished | 5.9 | 8,748 |
| Octopus and penguin band | 1 | Finished | 11.9 | 17,571 |
| Octopus and penguin band | 2 | Finished | 16.6 | 24,512 |
| Octopus and penguin band | 3 | Finished | 12.8 | 18,919 |
| Roast | 1 | Finished | 1.3 | 1,912 |
| Roast | 2 | Finished | 1.8 | 2,699 |
| Roast | 3 | Finished | 2.0 | 2,926 |
| Text adventure | 1 | Finished | 32.5 | 47,058 |
| Text adventure | 2 | Finished | 22.7 | 32,707 |
| Text adventure | 3 | Finished | 24.8 | 35,891 |
Build and settings
What exactly ran
- Maker
- Tencent
- Build
- AngelSlim SPARK-4Bit-Fast GGUF
- File size
- 149 GB
- Machine
- NVIDIA DGX Spark, 128 GB
- Runtime
- llama.cpp 2cbca38 with the Hy4 Spark patch
- Sampling
- temperature 0.9, top-p 1.0 (maker settings)
- Thinking
- on (default)
Speed
How fast it ran here
- Writing speed, short prompt
- 25.0 tok/s
- Writing speed after a 32K prompt
- 23.6 tok/s
- Writing speed after a 127K prompt
- 21.7 tok/s
- Reading speed, 32K prompt
- 675 tok/s
- Reading speed, 127K prompt
- 580 tok/s
- Wait for first token, 32K prompt
- 47 s
- Wait for first token, 127K prompt
- 219 s (3.6 min)
- Load time
- 160 s (2.7 min)
- Least memory left free
- 26 GiB
- Time for all 27 attempts
- 5.1 hours
- Tokens generated, all attempts
- 447,477 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.