The exact Hugging Face repository was not recorded in the approved run.
Source: MESSUNG_2026-10-13.json, recorded model-file size. Retrieved . Local model-file size recorded on 13 October 2026, rounded to 0.01 GB; exact bytes and Hugging Face repository were not recorded.
Hugging Face listings can change. These are dated file-size records; a revision hash was not retained. Download size does not prove that a runtime can load the model.
Fits on paper
Subtract this variant's recorded file size from each memory budget; the remainder is the upper bound for context and other overhead.
Decimal: 1 GB = 1,000,000,000 bytes; 1 TB = 1,000 GB. These are arithmetic comparison budgets, not verified usable hardware capacities. The answer can change with GB/GiB: 512 GiB = 549,755,813,888 bytes. The measured RTX 4080 has 16 GB VRAM; every other device name only labels a comparison budget.
16 GB · RTX 4080
files fit on paper (rounded size)
about 9.45 GB left
My measured card; usable free memory is not verified.
24 GB · RTX 3090/4090
files fit on paper (rounded size)
about 17.45 GB left
Comparison budget; not measured.
32 GB · RTX 5090
files fit on paper (rounded size)
about 25.45 GB left
Comparison budget; not measured.
128 GB · DGX Spark
files fit on paper (rounded size)
about 121.45 GB left
Comparison budget; I do not own one. Usable free memory is lower and not verified.
256 GB · Mac Studio M5 Ultra or 2× DGX Spark
files fit on paper (rounded size)
about 249.45 GB left
One Mac Studio M5 Ultra, or the sum of two DGX Sparks (two memories, not one pool). Not measured.
512 GB · Mac Studio M5 Ultra
files fit on paper (rounded size)
about 505.45 GB left
Planned; not delivered.
1 TB · 2× Mac Studio
files fit on paper (rounded size)
about 993.45 GB left
Across two planned machines; not one shared memory pool.
GB values are rounded to 0.01 GB for display; “left” is what remains for context, runtime and the operating system. Exact bytes are in this model's data (JSON).
Budget minus model files leaves an upper bound for context, runtime buffers and the operating system; no context length or runtime compatibility is promised. Exactly filling a budget leaves no room for context, so the answer is “no room for context on paper”; files larger than the budget are “files too big”.
Two Mac Studio machines, 512 GB each; not delivered. Two DGX Sparks are likewise two memories, not one pool. A model split across two machines needs a compatible runtime and must fit each machine's allocation. This table is arithmetic, not a cluster test.
Append-only run history · oldest first
Approved measurements
NVIDIA GeForce RTX 4080 · LM Studio version: not recorded.
Recorded medians by prompt length; input tokens are not the configured context window.
Prompt
Input tokens
Output tokens/s
First token (engine) (s)
Power (W)
Electricity per 1M tokens
Short
32
92.7
0.08
155.5
€0.16
Medium
855.5
94.5
0.3
163.6
€0.17
Long
3,432.5
93.3
0.74
174.5
€0.18
Power on 13 October was measured with a smoothed sensor and is too low; see the 27 October measurement using instantaneous power.
Recorded model-level electricity cost per 1M output tokens (short prompt): €0.16. Power scope: GPU only, €0.35/kWh. GPU power only; not whole-system power. Excludes hardware cost; not an API price.
Method and limits
Temperature: not measured yet; maximum output: 300 tokens; repeats per prompt length: 2. Fresh identifier and topic per request; no prompt cache. The source reports medians and individual output-speed samples.
Available exact commands and prompts are public in the raw-data record below. Exact prompts were not recorded for this run. Timezone was not recorded; this entry uses its local clock for ordering. Its absolute order against other timezones is unknown. Missing software versions and context limits remain missing. A small sample describes this run, not a reproducible benchmark or a speed guarantee.
Public extract of the approved MESSUNG file. It includes recorded individual values and available GPU sensor samples. Internal paths and operational notes are excluded. Missing records stay null.
Software
LM Studio · version not recorded
Driver version
not recorded
Model-file SHA256
not recorded
Exact recorded command
Not recorded. No command reconstructed from today's scripts.
Exact prompts by repetition
Not recorded. Task: text generation with fresh short, medium and long inputs; a new identifier and topic per request.
This record cannot yet reproduce the original run exactly. Missing: exact command not recorded; exact kurz prompts not recorded for every repetition; exact mittel prompts not recorded for every repetition; exact lang prompts not recorded for every repetition; software version not recorded; driver version not recorded; model-file SHA256 not recorded.
Recorded medians by prompt length; input tokens are not the configured context window.
Prompt
Input tokens
Output tokens/s
First token (wall) (s)
Power (W)
Electricity per 1M tokens
Short
35
93.1
2.13 · engine: 0.09 s
234.8
€0.25
Medium
868.5
93.2
2.13 · engine: 0.33 s
240
€0.25
Long
3,442
91.9
2.19 · engine: 0.7 s
242.1
€0.26
All generated tokens incl. reasoning. Wall TTFT is shown first; engine time is separate. Wall TTFT was derived from request duration minus generation time (non-streaming). Power sensor: power.draw.instant.
Recorded model-level electricity cost per 1M output tokens (short prompt): €0.25. Power scope: GPU only, €0.35/kWh. GPU power only; not whole-system power. Excludes hardware cost; not an API price.
Method and limits
Temperature: not measured yet; maximum output: 1,500 tokens; repeats per prompt length: 2. Fresh identifier and topic per request; no prompt cache. The source reports medians and individual output-speed samples.
Available exact commands and prompts are public in the raw-data record below. Timezone is included in the timestamp. A small sample describes this run, not a reproducible benchmark or a speed guarantee.
Public extract of the approved MESSUNG file. It includes recorded individual values and available GPU sensor samples. Internal paths and operational notes are excluded. Missing records stay null.
Software
LM Studio · 2.61.0 (engine version)
Driver version
not recorded
Model-file SHA256
not recorded
Exact recorded command
Not recorded. No command reconstructed from today's scripts.
Exact prompts by repetition
Not recorded. Task: text generation with fresh short, medium and long inputs; a new identifier and topic per request.
This record cannot yet reproduce the original run exactly. Missing: exact command not recorded; exact kurz prompts not recorded for every repetition; exact mittel prompts not recorded for every repetition; exact lang prompts not recorded for every repetition; driver version not recorded; model-file SHA256 not recorded.
The JSON is an extract of approved measurement fields. Internal paths and operational notes are excluded. New runs append here; the model URL stays the same.
Steven Kramp