Official FP8 child artifact on TP4/EP4, eager text, and automatic KV. MTP0 uses vLLM 658965050 plus kernels 2f829747; MTP1-4 use vLLM 1372c62d plus staged kernels 2f829747. This is a coverage map across exact runtime identities, not a controlled depth curve. Fixed: revision=qwen38-flash-next · artifact_id=qwen38-flash-next-fp8-bcd9f01 · runtime_family=vLLM XPU · tp=4 · ep=4 · parallel_profile=TP=EP · graph_mode=off · kv=auto · modality=text.

◇ Speed check only0, no speculative decoding 0·MTP05.22 tok/sreport · evidence
◇ Speed check only1K, no speculative decoding 1K·MTP05.13 tok/sreport · evidence
◇ Speed check only2K, no speculative decoding 2K·MTP05.23 tok/sreport · evidence
◇ Speed check only4K, no speculative decoding 4K·MTP04.46 tok/sreport · evidence
◇ Speed check only8K, no speculative decoding 8K·MTP03.98 tok/sreport · evidence
◇ Speed check only0, + speculative decoding (depth 1) 0·MTP19.37 tok/sreport · evidence
◇ Speed check only4K, + speculative decoding (depth 1) 4K·MTP18.9 tok/sreport · evidence
◇ Speed check only0, + speculative decoding (depth 2) 0·MTP211.9 tok/sreport · evidence
◇ Speed check only4K, + speculative decoding (depth 2) 4K·MTP29.89 tok/sreport · evidence
◇ Speed check only0, + speculative decoding (depth 3) 0·MTP314.89 tok/sreport · evidence
◇ Speed check only4K, + speculative decoding (depth 3) 4K·MTP315.5 tok/sreport · evidence
◇ Speed check only0, + speculative decoding (depth 4) 0·MTP420.73 tok/sreport · evidence
⚠ Quarantined1K, + speculative decoding (depth 1) 1K·MTP1 — The exact 32-block configured-1536 identity passed source, runtime, four-rank, placement, cache, capacity, and health gates, then reached the unchanged worker-response deadline during its first 1K request. No output was returned; request two was not sent and the separate 2K boot did not run. Observed: 768 computed prompt tokens; no output.evidence
⚠ Quarantined2K, + speculative decoding (depth 1) 2K·MTP1 — The exact 32-block configured-3072 identity passed all startup gates and exposed 7,561 cache tokens. Its first exact-2K exchange had a zero-byte completion body and no output token recorded when the fixed 360-second client bound expired; the subsequent engine diagnostic showed 448 computed prompt tokens and zero output, while vLLM completed-request, token, and MTP counters remained zero. The engine independently reported its own sampling timeout. Request two was not sent. The post-failure teardown window recorded compute- and copy-class resets on all four cards before all four were rediscovered; no post-reset collective was run. Existing MTP1 configured-512 and exact-4K passes and every captured speed remain unchanged. Observed: 360-second client timeout; 448 computed tokens; no output or speed.evidence
⚠ Quarantined8K, + speculative decoding (depth 1) 8K·MTP1 — The exact 32-block current-source identity exposed 13,516 cache tokens and completed one p8192/o128 request with exact usage, zero cache reuse, all 25 generic depth gates, and positive MTP1 counters at position zero. Its token array first diverged from the frozen cross-runtime/cache MTP0 authority at zero-based generated-token index 72. The 4.151 tok/s rate and 953.3-second TTFT receive no speed or quality credit. Controlled teardown passed its cleanup gates and returned all four cards idle; the intentional stop retained the known shutdown-time output-handler notice and one shared-memory cleanup warning. The host window contained corrected local-NVMe events but no B70 event. Observed: exact 8K completed; parity mismatch at token 73; 4.151 tok/s diagnostic only.evidence
⚠ Quarantined1K, + speculative decoding (depth 2) 1K·MTP2 — Both exact active-1K requests returned the frozen MTP0 text hash with zero cache reuse, identical text, and perfect MTP2 acceptance at positions zero and one. Request one observed 10.683 tok/s after first text and the repeat sentinel 12.642 tok/s. Eleven corrected local-NVMe events after the preregistered journal cutoff failed the strict clean-host gate, so neither rate receives speed, quality, or deployment credit. No event named a B70 address; existing MTP2 configured-512, active-2K, and exact-4K results remain unchanged. Observed: exact parity twice; corrected local-NVMe events; diagnostic only.evidence
⚠ Quarantined2K, + speculative decoding (depth 2) 2K·MTP2 — The exact 32-block configured-3072 identity passed all startup gates. Its first exact-2K request returned 128 tokens with zero cache reuse and active MTP2 counters, but the token array diverged from the frozen MTP0 authority at zero-based generated-token index 12. Request two was not sent. The observed 4.527 tok/s is diagnostic only and does not alter the passing MTP2 512 or exact-4K results. Observed: completed; target-parity mismatch at token 13; no speed credit.evidence
⚠ Quarantined8K, + speculative decoding (depth 2) 8K·MTP2 — The exact 32-block current-source identity exposed 11,264 cache tokens and completed one p8192/o128 request with exact usage, zero cache reuse, all generic depth gates, and positive MTP2 counters at both positions. Its token array first diverged from the frozen cross-runtime/cache MTP0 authority at zero-based generated-token index 26. The 6.235 tok/s rate and 649.7-second TTFT receive no speed or quality credit. The bounded host window contained corrected storage/root-port events but no B70 event; teardown was clean and all four cards returned idle. Observed: exact 8K completed; parity mismatch at token 27; 6.235 tok/s diagnostic only.evidence
⚠ Quarantined1K, + speculative decoding (depth 3) 1K·MTP3 — The exact 25-block local-NVMe identity passed source, runtime, fresh four-rank, placement, cache, capacity, served-identity, and health gates. Request one began under the frozen protocol, then the server received an external SIGTERM before the response completed. No request JSON, usage, output hash, or performance result exists. Partial server metrics showed six drafted and six accepted tokens with 1.000 acceptance at all three MTP3 positions, but those counters receive no parity, speed, quality, or deployment credit. Request two was not sent. Existing MTP3 configured-512, active-2K, and exact-4K results remain unchanged. Observed: external SIGTERM during request one; six accepted draft tokens; no completed response or speed.evidence
⚠ Quarantined2K, + speculative decoding (depth 3) 2K·MTP3 — The exact 25-block configured-3072 identity passed source, runtime, four-rank, placement, cache, capacity, and health gates. Its first exact-2K request returned 128 tokens with zero cache reuse and active MTP3 counters, but the token array diverged from the frozen MTP0 authority at zero-based generated-token index 4. Request two was not sent. The observed 5.932 tok/s is diagnostic only and does not alter the passing MTP3 512 or exact-4K results. Observed: completed; target-parity mismatch at token five; no speed credit.evidence
⚠ Quarantined8K, + speculative decoding (depth 3) 8K·MTP3 — The exact current-source TP4/EP4 eager MTP3 boot passed source/runtime, fresh four-rank, placement, served-identity, health, and 32-block capacity gates, reporting 9,654 cache tokens. Its sole p8192/o128 request reached the fixed 900-second client bound without a completed response receipt or any durably recorded output token. No usage, cache-zero, MTP-counter, parity, quality, TTFT, or speed result is claimed. The failed-request supervisor path cleanly removed the listener, process group, compile/RPC paths and returned all four cards idle. The host window contained corrected local-NVMe events but no B70-addressed event. Existing MTP3 configured-512, active-2K, exact-4K, and every captured speed remain unchanged. Observed: 900-second client timeout; no completed receipt or speed.evidence
⚠ Quarantined1K, + speculative decoding (depth 4) 1K·MTP4 — Both exact active-1K requests returned the frozen MTP0 text hash with zero cache reuse, identical text, and perfect MTP4 acceptance at positions zero through three. Request one observed 13.326 tok/s after first text and the repeat sentinel 17.291 tok/s. The exact stop sentinel terminated the timeout supervisor without reaching the detached server group; direct recovery produced an orderly shutdown, but the frozen teardown rule quarantines the cell. Seven corrected-only local-NVMe records separately block clean-host qualification. Existing MTP4 configured-512 and exact-4K results remain unchanged. Observed: exact parity twice; detached teardown gate failed; diagnostic only.evidence
⚠ Quarantined2K, + speculative decoding (depth 4) 2K·MTP4 — The exact 29-block configured-3072 identity passed source, runtime, fresh four-rank, placement, cache, capacity, served-identity, and health gates. Request one reached the fixed 360-second client bound without a response receipt; about five seconds later the engine independently reported its own sampling timeout at 384 computed prompt tokens and zero output. No HTTP status, usage, parity hash, MTP counter delta, or speed exists, and request two was not sent. The corrected supervisor returned zero and left no listener, recorded process, compile path, or RPC path, but the teardown window recorded compute- and copy-class resets on all four B70s. All four cards were rediscovered at low memory use; no post-reset collective or known-good generation canary was run. Existing MTP4 configured-512, active-1K, exact-4K, and all captured speeds remain unchanged. Observed: 360-second client timeout; 384 computed tokens; no output; four-card teardown resets.evidence
⚠ Quarantined4K, + speculative decoding (depth 4) 4K·MTP4 — The exact 29-block configuration admitted 4,352 tokens but stopped during the 4K quality request; no durable quality or timing result was produced. Observed: worker timeout at 3,904 computed tokens; no speed.evidence
⚠ Quarantined8K, + speculative decoding (depth 4) 8K·MTP4 — The exact 36-block current-source identity exposed 9,504 cache tokens and completed one p8192/o128 request with exact usage, zero cache reuse, all 25 generic depth gates, and positive MTP4 counters at all four positions. Its token array first diverged from the frozen cross-runtime/cache MTP0 authority at zero-based generated-token index 26. The 4.026 tok/s rate and 918.4-second TTFT receive no speed or quality credit. Controlled teardown passed its cleanup gates and returned all four cards idle; the five-second grace expired before the remaining EngineCore and workers were stopped, and the intentional stop retained the known output-handler and shared-memory cleanup notices. Six corrected local-NVMe endpoint records block clean-host wording, but no B70 event appears. The superseded attempt-1 command-identity stop sent no request and grants no matrix credit. Observed: exact 8K completed; parity mismatch at token 27; 4.026 tok/s diagnostic only.evidence

Full matrix and exact selectors

Archived MTP0 context anchors and estimates

Legacy vLLM 658965050 TP4/EP4 eager-text MTP0 only. The 4K/8K cells are formal exact-depth measurements; 24K/32K are deterministic Grade-D extrapolations with 50%-150% bands. They do not transfer to the latest runtime or qualify boot, fit, quality, deployment, records, or promotion. Fixed: revision=qwen38-flash-next · artifact_id=qwen38-flash-next-fp8-bcd9f01 · runtime=vLLM XPU 658965050 + kernels 2f829747 · runtime_family=vLLM XPU · tp=4 · ep=4 · parallel_profile=TP=EP · graph_mode=off · kv=auto · modality=text.

◇ Speed check only4K, no speculative decoding 4K·MTP04.46 tok/sreport · evidence
◇ Speed check only8K, no speculative decoding 8K·MTP03.98 tok/sreport · evidence
≈ Estimate24K, no speculative decoding 24K·MTP0 — qwen38-flash-next-mtp0-context-estimator 1.0.0≈ 3.33 tok/s (1.66–4.99)evidence
≈ Estimate32K, no speculative decoding 32K·MTP0 — qwen38-flash-next-mtp0-context-estimator 1.0.0≈ 3.17 tok/s (1.59–4.76)evidence

Full matrix and exact selectors

TP4 eager text deeper-context coverage

Official FP8 child artifact on current-source vLLM 1372c62d plus staged kernels 2f829747, TP4/EP4 eager text, and automatic KV. Single generic requests and bounded negatives are Grade-D matrix evidence until semantic and repeat qualification exists. Fixed: revision=qwen38-flash-next · artifact_id=qwen38-flash-next-fp8-bcd9f01 · runtime=vLLM XPU 1372c62d + staged kernels 2f829747 · runtime_family=vLLM XPU · tp=4 · ep=4 · parallel_profile=TP=EP · graph_mode=off · kv=auto · modality=text.

⚠ Quarantined16K, no speculative decoding 16K·MTP0 — The current-source 33-block identity first completed one generic p16384/o128 request at a diagnostic-only 5.219 tok/s. The semantic program then returned one correct 16,213-token fresh response, corrupted the identical same-server repeat into repeated text despite zero reported cache use, and stopped a separate fresh-server attempt at 1,600 computed prompt tokens with no first output after an RPC sampling timeout. Cleanup of the latter recorded eight B70 engine resets and 61 unsuccessful responses before all four cards re-enumerated idle. The cell is nondeterministically unstable and receives no speed, curve, quality, deployment, or headline credit; unchanged retries and 24K/32K serving remain blocked pending a material runtime treatment. Observed: 16K sometimes completes; same-server corruption and fresh-server timeout; diagnostic only.evidence
⚠ Quarantined16K, + speculative decoding (depth 2) 16K·MTP2 — The current-source MTP2 identity admitted exactly 40 blocks / 20,014 cache tokens in both bounded arms. Scheduler-64 stopped at 3,200 computed prompt tokens; the scheduler-32 arm was observed 2,240 tokens / 70% farther at 5,440, but the sole p16384/o128 treatment request still returned no output before the unchanged 300-second runtime response deadline. This is treatment evidence, not repeat-confirmed causality. Its shutdown was followed by eight card reset records and 58 unsuccessful card responses, so the frozen postflight rule fails and the treatment tranche is exhausted. No capability, speed, quality, parity, deployment, or headline credit is granted; current checks found no listener or owned residue and all cards are below 43 MiB. Observed: exact 16K stopped at 5,440 computed prompt tokens after scheduler-32 treatment; no output; four-card teardown events; diagnostic only.evidence

13 untested combinations

24K·MTP032K·MTP016K·MTP124K·MTP132K·MTP124K·MTP232K·MTP216K·MTP324K·MTP332K·MTP316K·MTP424K·MTP432K·MTP4

Full matrix and exact selectors

Card-fit summary

Official FP8 child artifact, vLLM 658965050 plus kernels 2f829747, eager MTP0 text, zero prior context, automatic KV, and TP=EP. TP1 and TP2 need separate fit/offload designs; they are not claimed unsupported. Fixed: revision=qwen38-flash-next · artifact_id=qwen38-flash-next-fp8-bcd9f01 · runtime=vLLM XPU 658965050 + kernels 2f829747 · runtime_family=vLLM XPU · parallel_profile=TP=EP · active_context_tokens=0 · graph_mode=off · kv=auto · modality=text.

◇ Speed check only4 cards, no speculative decoding TP4·MTP05.22 tok/sreport · evidence

2 untested combinations

TP1·MTP0TP2·MTP0

Full matrix and exact selectors

Graph and modality summary

Official FP8 child artifact, vLLM 1372c62d plus staged kernels 2f829747, TP4/EP4, MTP0, zero prior context, and automatic KV. Eager text is screened; PIECEWISE text has bounded compile-resource evidence but no API, quality, or speed result; vision remains unmeasured. Fixed: revision=qwen38-flash-next · artifact_id=qwen38-flash-next-fp8-bcd9f01 · runtime=vLLM XPU 1372c62d + staged kernels 2f829747 · runtime_family=vLLM XPU · tp=4 · ep=4 · parallel_profile=TP=EP · mtp=0 · active_context_tokens=0 · kv=auto.

◇ Speed check onlytext, eager text·eager5.22 tok/sreport · evidence
⚠ Quarantinedtext, graph text·graph — Attempt 7 loaded the complete checkpoint on all four ranks and entered PIECEWISE torch.compile with one compile thread. The phase-aware guard latched below its 30-GiB host-memory floor before the later TTM/global-OOM teardown window. Graph capture did not complete, the API never became healthy, and no client, output, quality, replay, or speed row exists. This is Grade-D bounded resource evidence only. Observed: all 131 shards loaded on four ranks; graph compile began; 30-GiB host-memory guard tripped; no API, quality, or speed.evidence

2 untested combinations

vision·eagervision·graph

Full matrix and exact selectors

Full 270-cell coverage contracts · 26 classified

Coverage · 2 matrices26/270 classified

16 quarantined244 missing9 screened1 closed

Text serving coverage

Exact text-serving combinations across topology, MTP depth, active context, and graph mode on the latest retained runtime. Older-runtime evidence is preserved in a separate archival contract and never silently transferred into these cells.

26/240

240 exact cells26 classified0 measured214 gaps240 retry-tagged

1 ■ closed9 ◇ screened214 · untested16 ⚠ quarantined

Fixed: runtime=vLLM XPU 1372c62d + staged kernels 2f829747 · runtime_family=vLLM XPU · parallel_profile=TP=EP

Break down by axis

Weight revisionqwen38-flash-next 26/240

Artifactqwen38-flash-next-fp8-bcd9f01 26/240

Tensor parallel1 1/802 0/804 25/80

MTP depth0 5/481 5/482 6/483 5/484 5/48

Active context0 7/301024 4/302048 4/304096 5/308192 4/3016384 2/3024576 0/3032768 0/30

Graph modeoff 25/120PIECEWISE 1/120

KV cacheauto 26/240

Modalitytext 26/240

Fixed-vision serving coverage

Vision capability and quality combinations under one versioned fixed-image fixture. Context depth is intentionally excluded until a vision anchor establishes meaningful image-plus-text token accounting.

0/30

30 exact cells0 classified0 measured30 gaps30 retry-tagged

30 · untested

Fixed: runtime=vLLM XPU 1372c62d + staged kernels 2f829747 · runtime_family=vLLM XPU · parallel_profile=TP=EP · active_context_tokens=0 · modality=vision · workload_profile=fixed-vision-fixture-v1-required

Break down by axis

Weight revisionqwen38-flash-next 0/30

Artifactqwen38-flash-next-fp8-bcd9f01 0/30

Tensor parallel1 0/102 0/104 0/10

MTP depth0 0/61 0/62 0/63 0/64 0/6

Graph modeoff 0/15PIECEWISE 0/15

KV cacheauto 0/30

Measured results

Every number links to its proof.

family data

Research speed screen by TP ◇ screened, experimental

Only TP4 has bounded screens; each MTP cell has three same-server samples under the declared workload, so no TP mini-curve is implied

official FP8 · eager MTP0official FP8 · eager MTP1official FP8 · eager MTP2 (variable screen)official FP8 · eager MTP3 (variable screen)official FP8 · eager MTP4missing

official FP8 · eager MTP0 4: 5.22
official FP8 · eager MTP1 4: 9.37
official FP8 · eager MTP2 (variable screen) 4: 11.9
official FP8 · eager MTP3 (variable screen) 4: 14.89
official FP8 · eager MTP4 4: 20.73

All measured metric summaries

Decode

TTFT

official FP8 · eager MTP0 4: 5,376
official FP8 · eager MTP1 4: 9,349
official FP8 · eager MTP2 (variable screen) 4: 11,278
official FP8 · eager MTP3 (variable screen) 4: 11,818
official FP8 · eager MTP4 4: 10,023

evidence 1 · evidence 2 · evidence 3 · evidence 4 · evidence 5

Legacy-comparable after-first-text context screen ◇ screened, experimental

The plotted 5.22/5.13/5.23/5.23 tok/s values use the legacy after-first-text comparison and stop at 4K. Separate formal cache-zero rates are 3.865 at 2K, 4.456 at 4K, and 3.980 tok/s at 8K; 8K is not plotted here because its third comparison row did not complete. All cells retain the short-quality caveat.

official FP8 · TP4 eager MTP0

official FP8 · TP4 eager MTP0 0: 5.22 · 1K: 5.13 · 2K: 5.23 · 4K: 5.23

All measured metric summaries

Decode

official FP8 · TP4 eager MTP0 0: 5.22 · 1K: 5.13 · 2K: 5.23 · 4K: 5.23

TTFT

official FP8 · TP4 eager MTP0 1K: 29,043 · 2K: 81,147 · 4K: 123,391

evidence 1

Exact-4K service tradeoff by MTP depth ◇ screened, experimental

Workload-aligned p4096/o256/c1 service screens with no harness-added warmups, but from different vLLM source revisions and cache allocations. MTP1 and MTP2 use 32-block headroom recipes; MTP3 remains faster with lower TTFT and higher wall output. Separate single-point series avoid implying a causal MTP-depth or cache curve; only MTP4 remains quarantined.

official FP8 · TP4 eager MTP0official FP8 · TP4 eager MTP1 · headroom32official FP8 · TP4 eager MTP2 · headroom32official FP8 · TP4 eager MTP3missing

official FP8 · TP4 eager MTP0 0: 5.23
official FP8 · TP4 eager MTP1 · headroom32 1: 8.9
official FP8 · TP4 eager MTP2 · headroom32 2: 9.89
official FP8 · TP4 eager MTP3 3: 15.5

All measured metric summaries

Decode

official FP8 · TP4 eager MTP0 0: 5.23
official FP8 · TP4 eager MTP1 · headroom32 1: 8.9
official FP8 · TP4 eager MTP2 · headroom32 2: 9.89
official FP8 · TP4 eager MTP3 3: 15.5

End-to-end output

official FP8 · TP4 eager MTP0 0: 1.49
official FP8 · TP4 eager MTP1 · headroom32 1: 0.98
official FP8 · TP4 eager MTP2 · headroom32 2: 0.89
official FP8 · TP4 eager MTP3 3: 1.25

TTFT

official FP8 · TP4 eager MTP0 0: 123,391
official FP8 · TP4 eager MTP1 · headroom32 1: 232,079
official FP8 · TP4 eager MTP2 · headroom32 2: 263,279
official FP8 · TP4 eager MTP3 3: 187,899

evidence 1 · evidence 2 · evidence 3 · evidence 4

Fine print

Qwen4Exp hybrid Gated DeltaNet/QSA sparse MoE, 48 layers, 2560 hidden. The FP8 repository is a quantized child of Qwen3.8 Flash-Next, not a separate model. Measurements remain artifact- and runtime-specific. Measurements, artifact hashes, outputs, quality decisions, and speed stay pinned to their exact recorded identity.

Many people at once Not measured

Multi-user (aggregate) throughput has not been measured for this family. The lab’s one measured sweep so far is in the multi-user report; the projection block below includes a projected users curve where a projection exists.

How much faster could this get? No projection

No projection is shown: this model is not in the ML Bottleneck catalog yet. The measured numbers above stand on their own.

Packets and recipes

Qwen3.8 Flash-Next FP8 · TP4+EP4 graph + lossless MTP1 + never-routed experts host-placed

Qwen3.8 Flash-Next FP8 · TP4+EP4 graph + lossless MTP1 + never-routed experts host-placed + Triton HC glue

Qwen3.8 Flash-Next FP8 · TP4+EP4 graph + lossless MTP1 + never-routed experts host-placed + both reference Triton kernels

Qwen3.8 Flash-Next FP8 ? TP4+EP4 graph, no speculation, W13-N64

Qwen3.8 Flash-Next FP8 · TP4+EP4 graph + lossless MTP1

Qwen3.8 Flash-Next FP8 · TP4 research server

Qwen3.8 Flash-Next FP8 · TP4 MTP0 current-runtime research quick check

Qwen3.8 Flash-Next FP8 · TP4 MTP1 research quick check

Qwen3.8 Flash-Next FP8 · TP4 MTP3 research quick check

Qwen3.8 Flash-Next FP8 · TP4 MTP2 research quick check

Qwen3.8 Flash-Next FP8 · TP4 MTP4 research quick check

What has been classified

Practical TP4 eager text coverage

Archived MTP0 context anchors and estimates

TP4 eager text deeper-context coverage

Card-fit summary

Graph and modality summary

Text serving coverage

Fixed-vision serving coverage

Measured results

Research speed screen by TP ◇ screened, experimental

Legacy-comparable after-first-text context screen ◇ screened, experimental

Exact-4K service tradeoff by MTP depth ◇ screened, experimental

Many people at once Not measured

How much faster could this get? No projection

Keep going