DavidAU TURBO & Swift vs Qwen3.8-27B: thinking tokens
Two fine-tunes that promise far fewer thinking tokens didn't deliver: TURBO thinks the most of all three (58% of its output), and Swift's claimed 58% reduction came out at 15%. One run per prompt would have told me the opposite story.
- Tested
- Setup
Radeon Pro W7800 48 GB — bare-metal workstationBare metal, Linux, ROCm (Ollama) · AMD Ryzen 5 7600 (6C/12T, AVX-512) · 64 GB DDR5-6400
- Runtime
- Ollama 0.34.0, Q8_0 + MTP baseline
- Models
qwen3.8:27b-q8_0-mtpQ8_0 + MTPbaseline
Qwen3.8-27B TURBOtested
Qwen3.8-27B Swifttested
Verdict
Neither claim holds up on my prompts. TURBO's answers are shorter, but it has the highest thinking share and is wildly inconsistent (±3,580 tokens on the same question). Swift thinks ~15% less, not 58%. My default stays the same.
Works
- Both fine-tunes answer ~15% faster per question than the baseline
- Swift keeps the baseline's generation speed (44.4 tok/s)
Doesn’t
- TURBO: 58% of its output is thinking — claimed 1/2 to 1/10
- TURBO: run-to-run spread up to ±3,580 tokens on one question
- Swift: 'up to 58% less thinking' measured as 15%
- Thinking share: baseline / TURBO / Swift
- 31 / 58 / 30%
- Seconds per answer
- 50.3 / 42.0 / 42.9
- Repetitions needed at temp 1
- ≥ 3
Tokens per answer
lower is better5 prompts × 3 repetitions, temperature 1
- Thinking
- Total
Data
| Thinking | Total | |
|---|---|---|
| Official Q8_0 + MTP | 648 tokens | 2,112 tokens |
| DavidAU TURBO | 938 tokens | 1,631 tokens |
| Swift | 553 tokens | 1,836 tokens |
Seconds per answer
lower is betterData
| Time | |
|---|---|
| Official Q8_0 + MTP | 50.3 s |
| DavidAU TURBO | 42.0 s |
| Swift | 42.9 s |
How I tested
The same five everyday prompts I use as my “thinkbench” — two coding tasks, a Turkish knowledge question, a Turkish creative task and a logic puzzle — each run three times at temperature 1, with thinking on. I measured total tokens per answer, thinking tokens, wall time and generation speed.
Results
| Official Q8_0 + MTP | DavidAU TURBO | Swift | |
|---|---|---|---|
| Tokens per answer | 2,112 | 1,631 | 1,836 |
| Thinking per answer | 648 (31%) | 938 (58%) | 553 (30%) |
| Seconds per answer | 50.3 | 42.0 | 42.9 |
| Generation speed | 43.8 tok/s | 38.7 tok/s | 44.4 tok/s |
What I found
- DavidAU’s “1/2 to 1/10 the thinking tokens” doesn’t hold up — its thinking share is the highest of the three. Its total is lower only because its final answer is short. It’s also all over the place: on one question, runs differed by as much as ±3,580 tokens.
- Swift’s “58% less thinking” is exaggerated — I measured a 15% reduction.
- A single run would have fooled me. In the first round, TURBO was the most frugal model; in the second, it thought the longest. At temperature 1, three repetitions is the bare minimum.
I deleted both fine-tunes after measuring them.