Radiance
FIAT LUX · FIAT TOKENS · GLORIA IN EXCELSIS DECODE
On the sixth day of October, on one rig, one boot and one harness, the engine of light was set against our deployed vLLM fork. BetterBench 0.6.0, quick mode, thinking ON.
✠ Conditions of the rite ✠
- Contender
- Radiance 1.0.13 · uncensored Flash-Next .rad
- Versus
- q38fn-gptq-kvram16u97 (vLLM 29.07.2)
- Cards
- TP 2 × R9700
- Memory
- DDR5 5000 MT/s · 192 GB
- Passes
- 5 passes per category
- Sampling
- temp 0.7
- Cache
- cold prefix cache
- Answered
- 290 of 290 requests answered, both engines
The Decoding of the Categories
Median decode tokens per second at batch 1, five passes each. Dashed lines mark each engine's weighted combined score.
The Multitude Gathered
Aggregate tokens per second with 1, 2, 4 and 8 requests in flight, 48 requests per level. Radiance rises with the crowd. The deployed engine stays flat, its per-request rate falling about as fast as the batch grows.
The Reading of the Scrolls
Prefill: prompt tokens ÷ time to first token, median of 8 passes, cold cache. Both engines read the identical scrolls. The 64k target is 47,056 tokens.
The First Word Spoken
The Confessional
- Quick mode is BetterBench's own "smoke check, not a publishable result": 5 passes per category and one run per engine, deployed engine first.
- The pre-registered rule called for confidence intervals that
betterbench runnever prints, so the formal verdict is "not separated". The individual runs don't overlap (code: 88–95 vs 141–200 t/s). - Engine and model format changed together, so this compares the two packages.
- Radiance's
--tp-wire wht6link between the cards is lossy above 25 rows, which covers prefill and batched decode. Single-stream decode is exact. It also keeps a 47.68 GiB n-gram table in pinned RAM. - Answer quality has not been tested yet. That is the next session.
Raw: evidence/raw/radtestS1-bb-{A,B}-20261006T005341Z.json · verdict: radtestS1-VERDICT-20261006T005341Z.md