Prefill ≠ generation
Prefill reads input. Generation writes output. A fast result in one stage cannot stand in for the other.
APPLE SILICON · OPEN BENCHMARK WATCH
How fast is the
M5 Ultra getting?
A living record of local AI inference—prefill, token generation, the kernels behind each move, and what the numbers could mean for selling tokens.
Latest sourced event: 28 Sep 2026 · Public log through 29 Sep 2026
TWO SEPARATE TESTS / 28 SEP 2026● ● ●
PREFILL · 32K4,759input tok/sFlash-Next · oQ6e · 80c · cache unknown
GENERATION · CONTEXT UNKNOWN158–165output tok/sFresh prompt · 1 stream · output unknown
32K PREFILL EVOLUTION3,579 → 4,759
27 SEP28 SEP
NO MATCHED PP/TG PAIR CLAIMEDVIEW SOURCE ↗
01Two metrics, never one vague speed number
02Context, quantization and concurrency stay visible
03Every data point links to original evidence
01 / PERFORMANCE
The progress, day by day.
The time axis runs earliest to latest. Every point still shows its PP, TG and test conditions.
FIRST VERIFIED ULTRA DAY21 SEP 2026
INFERENCE SPACE / ROTATABLE 3D
Drag to rotate
X / TIME · 2026/09/21 → 2026/09/29Y / PREFILL · Z / GENERATIONDrag the chart to rotate · tap a dot for details
OTHER MACS / WHAT-IF
Inco Splash, M5 Pro & M5 Max.
Measured source-Mac throughput and a separate, unverified M5 Ultra scaling range. Source context stays attached to every value.
倍率依據 / Scaling basis
M5 Pro 16c → Ultra 80c: PP ×2–4, TG ×1.5–2.5. M5 Max 40c → Ultra 80c: PP ×1.2–1.8, TG ×1.2–1.7. Apple lists 307 / 614 / 1,200 GB/s memory bandwidth and 16 / 40 / 80 GPU cores for these specific variants. These broad factors allow for less than linear software scaling; they are scenarios, not cross-device benchmark findings. Each projection preserves its own context and engine.
Apple Pro / Max specs ↗ · Apple Ultra specs ↗Official Splash context curve · M5 Pro 16c / 48GB
| Model / metric | Short | 8K bucket | 16K bucket | 32K bucket |
|---|---|---|---|---|
| 35B PP · uncached | — | 2,575 | 2,373 | 2,011 |
| 35B TG · 1 stream | 210 | 156 | 149 | 143 |
| 27B PP · uncached | — | 398 | 395 | 363 |
| 27B TG · 1 stream | 74 | 55 | 55 | 54 |
SPEED-Bench coding prompts, reasoning on (27B medium), DFlash 2, HTTP, output cap 1,024; actual output lengths and TG cache state unpublished. Four short-stream aggregate: 35B 357; 27B 170. Cached 32K replay TTFT is a separate test. Inco official results ↗
Weinbach prototype v13–v17 · M5 Ultra
| Metric / context | v13 | v14 | v15 | v16 | v17 |
|---|---|---|---|---|---|
| PP · 2K prompt | 4,014 | 4,097 | 4,196 | 4,224 | 4,180–4,300 |
| TG · 10 fresh, starting context unknown | 51.5 | 102.5 | 118.9 | 115.9 | 165.7 |
| TG · 2K coding, 1 stream | 62.0 | 75.0 | 124.2* | 85.9 | 179.4 |
| TG · 2 streams aggregate | — | — | — | 153.2 | 138.2 |
| TG · 4 streams aggregate | — | — | — | 220.3 | 159.1 |
All versions appeared in one post on 24 Sep (Taipei). Author confirmed MTP and said v17 batch is broken. Exact quantization, implementation, fresh-prompt context, output length, cache/reasoning and batch context were not published in the table. *v15 has an unexplained source footnote. Original table ↗ · Author reply ↗
02 / CHANGELOG
What actually changed.
Newest sourced results first; the oldest records are at the bottom. Each model and patch links to its source.
03 / TOKEN ECONOMICS
What could this earn?
A transparent capacity scenario, not observed demand or a promise of profit.
SELECT A RECIPE
Pricing & operating assumptions＋
AT YOUR ASSUMPTIONSMEASURED PAIR
Monthly operating cash——
Gross / day—
Gross / month—
Simple payback—
UTILIZATION SENSITIVITYNET / MONTH
Capacity assumes prefill and decoding consume the device sequentially. Real sales also depend on queueing, latency, caching, quality, uptime, distribution and demand.
04 / READ THE DATA
The conditions matter.
Keep recipes together
Compare model, chip, quantization, context, cache state, prompt and concurrency. Different recipes appear as separate points.
Commercial use needs its own check
Flash-Next's model license requires separate permission for Model as a Service. Price snapshots and theoretical capacity are not contracted revenue.
MODEL LICENSE ↗CAPACITY MODEL
output tok/s = 1 / (input:output ratio / PP + 1 / TG)A conservative sequential-time approximation. Multi-user serving should use measured aggregate throughput with a latency target.