CALCULATED EXAMPLE
24 GB NVIDIA GPU
Everyday chat · 8K context · one session
- Memory pool
- Dedicated VRAM
- Result
- Ranked exact artifacts
- Evidence
- Estimates stay labelled
MODEL × HARDWARERESEARCH BETA
Can I actually
run it?
A compatibility lab for local AI. We separate what fits on paper from what has actually survived a run.
3 owned-hardware results10 retained cloud attemptsreceipt-v2 evidenceFailures preserved
PLANNING TOOL
Start with your computer. Or your next model.
Explore NVIDIA GPUs, DGX Spark, Apple Silicon or system RAM. Compare exact local model files, see the memory arithmetic, and get commands when runtime support is verified.
Find models that fit your computerDIRECT HARDWARE ANSWERS
Start with a common memory target
OWNED-HARDWARE PILOT
Exactly three measured Local Lab results
Retained runs on one Ryzen 9 7950X / RTX 4090 / 128 GB system. Each card separates synthetic decode, context-fill memory, and streaming power so unlike measurements are not blended.
OWNED HARDWARE · REVIEW PASSObserved
Qwen3.8-27B
Ryzen 9 7950X · RTX 4090 24 GB · 128 GB RAM
- Runtime
- llama.cpp b10453 · CUDA · full GPU offload
- Quantization
- Q4_K_M
- Context
- 64K allocated · FP16 KV · passed
- Decode
- 49.09 tok/sMedian of five llama-bench 256-token synthetic decode repetitions
- Memory
- 20,266 MiB (19.79 GiB)Sampled GPU peak during selected 64K context-fill run
- Power
- 65.90 W median · 281.38 W peakMedian and peak across retained 8K streaming telemetry
Failure conditions
No runtime failure was observed in the retained 8K, 16K, 32K, or 64K FP16-KV context-fill runs. Contexts above 64K were not tested in this owned-hardware series.
OWNED HARDWARE · REVIEW PASSObserved
Qwen3.6-27B
Ryzen 9 7950X · RTX 4090 24 GB · 128 GB RAM
- Runtime
- llama.cpp b10453 · CUDA · full GPU offload
- Quantization
- Q4_K_M
- Context
- 64K allocated · FP16 KV · passed
- Decode
- 49.04 tok/sMedian of five llama-bench 256-token synthetic decode repetitions
- Memory
- 20,266 MiB (19.79 GiB)Sampled GPU peak during selected 64K context-fill run
- Power
- 64.24 W median · 337.72 W peakMedian and peak across retained 8K streaming telemetry
Failure conditions
No runtime failure was observed in the retained 8K, 16K, 32K, or 64K FP16-KV context-fill runs. Contexts above 64K were not tested in this owned-hardware series.
OWNED HARDWARE · REVIEW PASSObserved
Gemma 4 31B-it
Ryzen 9 7950X · RTX 4090 24 GB · 128 GB RAM
- Runtime
- llama.cpp b10453 · CUDA · full GPU offload
- Quantization
- Q4_K_M
- Context
- 64K allocated · Q8_0 KV · passed (FP16 KV failed)
- Decode
- 45.00 tok/sMedian of five llama-bench 256-token synthetic decode repetitions
- Memory
- 21,956 MiB (21.44 GiB)Sampled GPU peak during selected 64K context-fill run
- Power
- 63.79 W median · 333.83 W peakMedian and peak across retained 8K streaming telemetry
Failure conditions
64K with FP16 KV failed during context allocation: cudaMalloc could not allocate a 1,200 MiB KV-cache buffer. The same 64K workload passed after changing only KV cache to Q8_0; 32K FP16 KV also passed.
Gate: this pilot stops at three reviewed cards. Expand only after meaningful search impressions, citations, or subscriber demand.
SOURCE FRESHNESS
What is monitored—and what still needs review.
Source discovery checks official model revisions, pinned GGUF artifacts, llama.cpp releases and hardware source pages. It can create review work, but it cannot run benchmarks, spend money, alter an Observed label, deploy or publish.
model revisionQwen3.8-27B official revisionChecked 2026-09-20 22:29 UTC
Currentmodel revisionQwen3.6-27B official revisionChecked 2026-09-20 22:29 UTC
Currentmodel revisionQwen3-8B official revisionChecked 2026-09-20 22:29 UTC
Currentgguf artifactqwen3.8-27b Qwen3.8-27B-Q4_K_M.ggufChecked 2026-09-20 22:29 UTC
Currentruntime releasellama.cpp releasesChecked 2026-09-20 22:29 UTC
Review requiredhardware specificationNVIDIA GeForce RTX 4090 specificationsChecked 2026-09-20 22:29 UTC
Currenthardware specificationNVIDIA GeForce RTX 5090 specificationsChecked 2026-09-20 22:29 UTC
Currenthardware specificationMac Studio M4 Max · 64GB specificationsChecked 2026-09-20 22:29 UTC
Currenthardware specificationMac Studio M3 Ultra · 512GB specificationsChecked 2026-09-20 22:29 UTC
CurrentObservedRun receipt + raw measurements
VerifiedEvidence independently checked
ReportedNamed external source
EstimatedTransparent formula, not a test
OUR STANDARD
Every number needs a receipt.
Model revision. File hash. Runtime version. Context. KV cache. Peak memory. Prompt and generation speed. Failure boundary. If it cannot be reproduced, it is not an observed result.
- 01IdentifyImmutable model and hardware records
- 02RunWarm-up plus repeated measurements
- 03VerifySchema, hash and consistency checks
- 04PublishRaw receipt and limitations included
Evidence first. Estimates labelled. Corrections preserved.