CROWN CITADEL RESEARCH REPORT · 06 OCT 2026
Engine release ↗Strata Flashv0.1.40
Qwen3.8 Flash Next · UD-IQ4_XS · 128 GB Strix Halo
RECORDED BENCHMARK REPLAY
Flash Race
Benchmark clock
00:00.00
replay speed ×10
Speed
start0:00
passpartial creditfailin progress
Race feed
Live standings · points
THE RECORDED COMPARISON
Results at a glance
Strata is highlighted in ivory. Historical stacks keep their saved protocols; native clocks and cache histories differ. These plots do not isolate an engine-only gain.
Agent quality
Hermes: two complete 20-case passes. Tools: TC70–84 with each model’s recommended nonthinking sampler. ROCmFPX2: official 73, separate reviewed 80.
Agent case wall
Sum of scored case durations. Loading, transport probes, restoration and operator repair gaps are excluded. Speed does not imply quality.
Native decode · HE0–9
Thinking-off speed panel; 10/10 Strata outputs usable. Strata counts first prediction and terminal EOS; llama stacks count n−1; Gufo and Halogen count n. No correctness score or API native-timing estimate.
HE0–9 request wall
One selected run, 10 requests, own model-card sampler. Strata: 29.76 s. Saved runtime and serving differences remain in the source.
PROMPT PROCESSING
Prefill and context
Zero prefix replay: 18/18 append cellsNative prompt-read equals new input; reused prefix equals the requested depth. Counting output was clean in 2/18 cells. First deviations are preserved.
Cold prefill
Five complete inputs: 32K, 64K, 120K, 192K, 256K. Strata’s unscored primers supply the cold telemetry; the required warmup is separate.
Append prefill
Only new tokens enter Strata’s timed append. Historical plots preserve native PP and prior cache lifecycles, including any reported replay. Hover for replay counts.
Decode after append
512 output tokens per cell. Strata’s counting output was clean in 2/18 cells; its decode values are diagnostic where the first output deviated. Hollow markers identify those cells.
Time to first output
Observed client time in seconds, including new input processing. Identical frozen requests and append order; no retry of output deviations.
TARGET-ONLY · COMMON BF16 TEACHER
Full-vocabulary fidelity
16 frozen 512-token windows · last 128 positions · 2,048 rows · 248,320-token vocabulary.
KL divergence ↓ lower
Top-1 agreement ↑ higher
Perplexity ↓ lower
Strata target-only: KL 0.05976 · top-1 93.95% · PPL 2.0863. MTP disabled; the IQ-pack startup requires spec2. Binary logits are retained privately; public hashes and scorer source are included.
SERVING FOOTPRINT
Disk and memory
Serving assets
Strata: 89.41 GiB for weights, native pack and MTP, deduplicated by inode. Build/runtime caches excluded. Historical package scopes remain as recorded.
Recorded memory footprint
Hover for each method. Strata: 79.29 GiB peak MemAvailable drop from its pre-load baseline. Halogen reports pinned allocations. UMA host RAM, GTT, VRAM and RSS overlap.
| Strata panel | Ready RAM | Peak RAM | Minimum available |
|---|
Protocol, boundaries and full source
Active September 28 sweep plus the September 29 recommended-thinking Hermes correction. One selected run per panel; the selected protocol’s unscored warmup is retained. Hermes includes its required three-case smoke plus two scored 20-case passes. No new baseline or API reruns.
Strata v0.1.40 is pinned to 1735d6471df29b42c26170efaac1f1446a58640f. ROCm 7.14.1, gfx1151, context 262144, prefill 16384, int8 KV, resident experts, spec4, MTP Q4 all, lookup chain3. Optional numeric fast kernels are off; all 64 HIP compile commands use O3. Fidelity uses target-only spec2. Tool and Hermes execution audits match agent calls to completed environment events.
Native timer definitions differ. Strata’s PP counts the last input token while its work is charged to the decode clock. Historical server lifecycles and Carlos’s unrecovered lookup history limit controlled speed claims. Mismatched protocols and infrastructure startups are preserved as diagnostics. HE0–9 is a speed panel; full HumanEval, Core-19 and unrelated long-output tests are outside this scope.
The full source repository includes engine code, benchmark runners, frozen requests, model cards, audits, sanitized raw responses, timing/resource records, report source and checksums. Private identifiers and synthetic secrets are redacted; numeric results and first outcomes are preserved. Weights, runtime executables, environment homes and binary logits are excluded.