A faster challenger: Gyro-S
In short Gyro-S decodes 1.5 to 1.7 times faster than Ollama's 8B, restarts a known prompt in 0.065 s, and passes 71 of 81 journeys on the suite prompt (the 35B 72, the 8B 32). On the production prompt the 35B fast holds 72 and Gyro-S falls to 62, so the 35B fast is still the safer decision-maker. The speed is the best in the series; the open question is fit.Full verdict ↓
After the 35B, I ran a different kind of candidate: Gyro-S, a lighter pack of Qwen3.8-Flash-Next served by the same engine. If the answer to post 8's question is a small, fast model, this is the first test of it.
| Measured on the same harness | Ollama, 8B | llama.cpp, 8B | Strata, 35B | Strata, Gyro-S | |
|---|---|---|---|---|---|
| Decode at 1k | 101.3 tok/s | 109.2 tok/s | 54.1 tok/s | 153.7 tok/s | fastest |
| Decode at 8k | 89.0 tok/s | 93.0 tok/s | 48.5 tok/s | 154.8 tok/s | fastest |
| Cold first token at 1k | 0.356 s | 0.350 s | 0.754 s | 1.187 s | slowest |
| Cold first token at 8k | 1.879 s | 1.854 s | 4.581 s | 3.213 s | middle |
| Warm first token | 0.16 to 0.20 s | 0.16 to 0.18 s | equal to cold | 0.065 s | fastest |
Result 1: decode and the physics
Post 5's rule says generation cannot exceed bandwidth divided by the bytes read per token. At 154 tok/s on a 760 GB/s card, Gyro-S can read at most about 4.9 GB per token (760 ÷ 154). That is below the 5.0 GB of the 8B in post 5, so it either has fewer active weights or fewer bits per weight. The number is a ceiling on its size, not its size, and I do not yet have the real figure.
Decode also dipped at 4k (133 tok/s) between two runs near 154. I have not explained it, and it may be noise at three repetitions.
Result 2: a prompt it has seen
Post 6 found that Strata could not reuse a prompt, and that this made the 35B a poor fit for agent loops. Gyro-S does reuse it: the warm first token is 0.065 s at 1k, 4k and 8k alike, against 1.19 to 3.21 s cold. So the engine can do it. Whether the 35B pack cannot, or was simply run without it, I have not checked, and that changes what post 6 should say about Strata.
The cost is the cold read: Gyro-S takes 1.19 s for an 800-token prompt where Ollama takes 0.36 s. For an agent that mostly resends a long prompt, warm start is the number that matters. For one that starts fresh every time, it is the cold one.
Result 3: decisions
The corrected journey suite (27 scenarios, 3 samples) and the 15-prompt tool-calling check:
| Check | Gyro-S | 35B-A3B | 8B |
|---|---|---|---|
| Journeys, suite prompt, fast | 71/81 (88%) | 72/81 (89%) | 32/81 (40%) |
| Journeys, production prompt, fast | 62/81 (77%) | 72/81 (89%) | not run |
| Tool calls (15), fast | 12/15 | 15/15 | not run on this check |
Gyro-S is level with the 35B on the journey suite and far ahead of the 8B, but only on the suite's own prompt. With the production prompt the 35B fast holds 72/81 and Gyro-S fast falls to 62/81, losing the hotel preference, the tool-stays-down disclosure and the Rome trip. The 35B fast is still the safer decision-maker under the prompt the agent actually uses; Gyro-S's case is speed and a possible fit.
Thinking is a trade. It lifts Gyro-S's tool-call score from 12/15 to 14/15, but the journey score barely moves: 69/81 against 71 with the suite prompt, and 66/81 against 62 with the production prompt at 25.7 s a case. Without thinking it picked the wrong month (2025-10) on the grocery questions and missed the second call in a comparison; with thinking it fixed those but failed the "Send an email to Sam" restraint case. The 35B is 15/15 without thinking and 14/15 with it, where thinking drops the parallel reminder.
The safety caveat is the 35B's, not Gyro-S's: the 35B thinking mode follows the held-out page's instruction to call trip_book in six of six samples across both prompts, while fast mode, Gyro-S and the 8B refuse. The quality totals hide that.
Gotchas
| Gotcha | Effect |
|---|---|
| One request in flight measured 6.5 tok/s | Two requests measured 65 and four 80. The single figure is about 12 times lower than the others would suggest and I believe it is a stall in that run, not the model. I have not rerun it, so do not quote it |
| Aggregate at four requests (80 tok/s) is below single-stream decode (154) | The throughput test includes reading fresh 1k prompts, which are cold every time, so it measures prompt reading as much as generation |
| Size and quantization unknown | The ceiling above bounds the model, but I cannot say what fits beside the other services |
| Fit not measured | It ran with the rest of the stack paused, as the 35B did. Whether it can coexist is the next experiment |
Where it stands, and my verdict
The setups against the criteria from post 1:
| Setup | Decode at 1k | Warm first token | Decisions | Resident beside the stack |
|---|---|---|---|---|
| Qwen3 8B, Ollama | 101 tok/s | 0.16 s | 32/81 journeys | yes |
| Qwen3 8B, llama.cpp | 109 tok/s | 0.16 s | not run | yes |
| 35B, Strata | 54 tok/s | 0.75 s | 72/81 journeys | no |
| Gyro-S, Strata | 154 tok/s | 0.065 s | 71/81 journeys; 62/81 with the production prompt | not measured |
The decision column uses the corrected journey suite (27 scenarios, 3 samples), not the old decision suite; post 7 has the update.
Verdict: the best speed result of the series, but not yet a quality win. Gyro-S beats the 8B on decode and on warm start, and the engine can reuse a prompt after all. It passes 71 of 81 journeys on the suite prompt, level with the 35B's 72 and far ahead of the 8B's 32. On the production prompt the 35B fast holds 72 and Gyro-S falls to 62, so the 35B fast is still the safer decision-maker. Gyro-S's case is speed and a possible fit, not a quality win; the next experiment is whether it can sit beside the stack.