llmprobe report card

/Volumes/edata2/models/afm/DeepSeek-V4-Flash-0731-AFM-MLX-native-v11

http://127.0.0.1:10304/v1

Overview

Three independent scores — never averaged

Surface coverage
100%
Core 9/9 core complete
core 100%
extended 80%
frontier 88.9%
How much of the standard API surface exists. Missing features are listed on purpose.
Engine conformance
100%
192/192 MUST no MUST fails
Of the surfaces that exist, how correct are the MUST behaviors. Unsupported ≠ fail.
Model capability
100%
strong 8 categories
Practical floor for tools, JSON, instructions — graded below floor / capable / strong.
Performance
29.1
tok/s decode 369 ms first token
Informational — hardware-dependent and never scored. Same-machine comparisons only.
Agentic
3/3
Harder multi-step bar. Never blended into capability.
Engine fidelity
100%
Same-model only — holds the model constant so the number is the engine.
Outcomes honesty
Pass88
Fail0
Unsupported8
Inconclusive3
Skipped15
Unsupported and inconclusive are not zeros and not fails.

Surface coverage engine

Core 100%

Per tier, never averaged. Click Core / Extended / Frontier to expand every feature under that tier.

Click a tier row to expand · missing features sort first

✗ vision✗ reasoning items✗ reasoning history round-trip
✗ MCP tools
not probed: prompt caching
○ sampling extensions: top_k honored; min_p, repetition_penalty accepted (acceptance only) — detected, not scored

Engine conformance engine

100%

MUST assertions on implemented surfaces only. Click a surface tile to filter the table. Default view: failures only.

Unsupported and inconclusive are not failures — they do not enter the conformance denominator. Use the filters below to inspect every check.

Test Surface Assertion Evidence Status

Model capability model

100% · strong

Floor check — not an intelligence rank. Category floor is 50%. Click a category to expand its evals (failures first).

Click a category row to expand evals

Agentic model

3/3 tasks

Multi-step tool use in a simulated workspace — harder than the capability floor, never blended into it.

reads the config instead of answering from priors
3 steps
finds where the port really lives and edits only that
4 steps
follows the pointer in build.cfg instead of guessing
4 steps

Engine fidelity engine

100%

Same-model comparisons only. Click a slice to see what was measured. Unmeasured slices are named — never zeroed.

Performance engine

29.1 tok/s

Informational — never scored and never part of the exit code. Hardware-dependent: only comparable against runs on the same machine.

machine: Apple M3 Ultra · 512 GB · darwin arm64
Decode throughput
29.1
tok/s · median of 3
Time to first token
369
ms · median of 3
Prefill throughput
285.6
tok/s · median of 3
MeasurementResultHow it was taken
Decode throughput29.1 tok/s 29.1–29.1 over 3steady-state generation while writing code
Time to first token369 ms 368–369 over 3latency before the first generated token
Prefill throughput285.6 tok/s 285.5–285.9 over 3measured on a 1797-token prompt
SignalVerdictDetail
Prefix cachefail not detected1× — 5s cold → 5s warm; usage reports 0 of 1418 prompt tokens cached
Concurrency (4 streams)partial serialized0.24 efficiency · 23.3 tok/s aggregate vs 24 alone · slowest first token 12.6s
Speculative decodepass 1.82× effectivepredictable 46.1 tok/s vs novel 25.3 tok/s · 4.83 tokens per decode step
Sustained loadpass steady29.1 → 29.1 tok/s over 390s (-0.1%)

Context scaling — decode, first-token latency and speculation as the prompt grows. A failed rung ends the ladder; larger sizes were not attempted.

Decode vs context tok/s 011213243 5074.1k8.2k16.3k decode · 507 tok → 38.3 tok/sdecode · 4.1k tok → 31.9 tok/sdecode · 8.2k tok → 29.1 tok/sdecode · 16.3k tok → 19.9 tok/s First token vs context ms 019.6k39.2k58.7k78.3k 5074.1k8.2k16.3k ttft · 507 tok → 1829 msttft · 4.1k tok → 15572 msttft · 8.2k tok → 32352 msttft · 16.3k tok → 69919 ms
PromptDecodeFirst tokenPrefillSpeculation
~507 38.3 tok/s 1.8s 277 tok/s 3.62 tok/step · 1.34× ceiling
~4.1k 31.9 tok/s 15.6s 265 tok/s 2.82 tok/step · 1.43× ceiling
~8.2k 29.1 tok/s 32.4s 254 tok/s 4.86 tok/step · 1.1× ceiling
~16.3k 19.9 tok/s 69.9s 234 tok/s 3.76 tok/step · 1.39× ceiling