Published measurements use one NVIDIA GeForce RTX 5090 through Infernix's public HTTP serving route. Choose a model below for its detailed results, run conditions, output limitations, and reproduction commands. These are recorded measurements; a model/backend being supported does not mean every workload or concurrency has a published measurement.
Read the measurement and publication rules for workload definitions, metric formulas, statistics, comparison requirements, and the standard result-page format.
Each cell links to the relevant result section. “Not published” describes measurement coverage, not product support. C is configured request concurrency; K is the number of draft tokens.
| Model / weights | MTP0 context profile | Single-request speculative decode | Corpus makespan | MTP3 decode saturation |
|---|---|---|---|---|
Qwen3.6-27B / groupwise-int |
8K–256K | MTP3 | Not published | C=1, 2, 4, 8 |
Qwen3.6-27B / nvfp4 |
8K–256K | MTP3 | Not published | C=1, 2, 4, 8 |
Qwen3.6-35B-A3B / groupwise-int |
8K–256K | MTP3; DFlash K=7 stochastic/greedy | MTP3 C=1, 2, 4, 8; DFlash C=1 | C=1, 2, 4, 8 |
Qwen3.8-27B / groupwise-int |
8K–256K | MTP3; DFlash2 K=7 | MTP3 C=1, 2, 4, 8; DFlash2 C=1 | C=1, 2, 4, 8 |
Qwen3.8-27B / nvfp4 |
8K–256K | MTP3; DFlash2 K=7 | MTP3 C=1, 2, 4, 8; DFlash2 C=1 | C=1, 2, 4, 8 |
Each page dates its runs (the Qwen3.6-35B-A3B page 2026-09-07, the Qwen3.8-27B page 2026-09-28/29, the Qwen3.6-27B page per run); later releases changed these models' decode routes, so the figures describe those runs' builds, not the current one.
Qwen3.8-Flash-Next has no row: its published figures are the README's comparisons with Strata, strata-nvfp4 and Basalt and Strata and Basalt on IQ3_S, measured with that section's client-side method rather than this page's, and its own guide's performance notes.
Qwen3.8 and Qwen3.6-35B-A3B C=1 corpus points also supply their single-request phase tables. The Qwen3.6-27B NVFP4 MTP3 phase table comes from a corpus C=1 point whose full makespan is not published here. Qwen3.8 measurements use FP8 E4M3 row-256 KV; the Qwen3.6 measurements use INT8 group-64 KV. Each model page records its build and run conditions.
| Question | Metric to use |
|---|---|
| How fast is prompt processing or an individual decode phase? | Prefill phase, Server TTFT, Decode phase |
| How long does the full fixed request set take? | Corpus makespan, Corpus decode, Requests/s |
| What aggregate decode rate is sustained at a full batch? | Steady decode |
These rates use different time boundaries. Server TTFT is the server's logged first-token time
(request_done.timings_seconds.ttft: preparation, queue wait and every prompt phase; corpus result
schema 9, which does not mix with the earlier phase sum); external
streaming TTFT has its own benchmark contract. Stochastic runs
can generate different token totals even with the same prompts and seeds. Output-limit and
repetition samples remain labeled in the measured corpus; throughput alone does not establish
successful task completion. See the 35B termination and anomalies
and Qwen3.8 completion outcomes.
- Serving benchmark runners: usage and local report files.
- Engine and Op benchmarks: their separate measurement scopes and commands.
- Capability evaluation: evaluation workflow; published scores live in the model cards, with the README's benchmark summary.
- Perplexity: offline causal-scoring measurement and comparison rules.
Model pages are the detailed result authority. README and model-card performance tables are excerpts linked to those pages; update them together when replacing an applicable measurement.