This week’s practical question: how does the hardware decision change as you move from a 24B or 27B model to a 120B-class MoE? The answer is not “buy the card with the biggest number.” Each step crosses a different artifact, context, runtime and platform boundary.
Three model classes, three buying decisions
| Workload class | Evidence boundary | Practical starting path |
|---|---|---|
| 24B LLM | 14.3GB Q4, 16.8GB Q5, 25.1GB Q8, 47.2–48.023GB BF16 | 24GB for Q4/Q5; 32GB for Q8; 64GB+ for BF16 planning |
| 27B LLM | Qwen3.8 community GGUF: 13.146GB–29.047GB; official FP8: 30.867GB; official standard: 55.563GB | 16GB constrained low-bit; 24GB practical Q4/Q5 start; 32GB quantized headroom; 48–64GB FP8; 80GB standard |
| gpt-oss-120b | OpenAI: at least 60GB in the cited MXFP4 path or multi-GPU; fits on one H100 | Bounded 64GB, H100 80GB, 96GB, unified-memory or deliberate multi-GPU paths |
What changes between 24B and 27B
The parameter difference looks small, but the exact artifacts change the decision. The listed 24B Q5 file is 16.8GB, while Google’s official Gemma 3 27B QAT Q4_0 model is 17.2GB before its separate 858MB vision projector. Both exceed a nominal 16GB pool. A single 24GB GPU is therefore the practical common starting tier, but context, vision, cache and runtime state decide the remaining margin.
At Q8, the cited 24B file is 25.1GB and the cited Qwen3.5-27B Q8-plus-projector pair is 29.523GB. Both point beyond 24GB; the 27B pair leaves only 2.477GB nominally on a 32GB device before runtime allocations. This is why 24GB, 32GB and 48GB are workload boundaries rather than marketing tiers.
Qwen3.8 updates the 27B decision
The current Qwen3.8-27B evidence separates Qwen’s official checkpoints from a separately maintained community GGUF conversion. The community files span 13.146GB at UD-Q3_K_XL, 16.464GB at UD-Q4_K_M, 19.772GB at UD-Q5_K_M, 23.088GB at UD-Q6_K_M and 29.047GB at Q8_0. The optional F16 vision projector (0.928GB) and Q4_0 MTP draft file (1.370GB) are separate allocations rather than hidden inside those model sizes.
That makes 24GB the practical consumer starting tier for community Q4/Q5 artifacts and 32GB the roomier quantized tier. It does not make 24GB an official-checkpoint target: Qwen’s official FP8 files total 30.867GB, while the standard checkpoint totals 55.563GB. Read the full Qwen3.8-27B memory ladder, then follow the text-first llama.cpp or Ollama workflow. On Apple silicon, use the shared-memory Mac guide before choosing a capacity tier. If you are still choosing hardware, stage the private Qwen3.8 24GB/32GB watchlist preset; it does not save or transmit anything until you explicitly choose Save.
What changes at 120B
gpt-oss-120b is a mixture-of-experts model with 117B total and 5.1B active parameters. The active count describes routing per token; it does not make the stored model behave like a dense 5.1B model. OpenAI’s published MXFP4 guidance—at least 60GB or multi-GPU—places it in a different hardware class.
- 64GB is bounded: only 4GB above the published floor before cache and runtime state.
- H100 80GB is the official single-GPU class: exact PCIe/SXM/NVL identity and host still matter.
- 96GB and 128GB paths need software evidence: RTX PRO 6000, DGX Spark and large-memory Macs are not interchangeable runtimes.
- Multiple GPUs are not one pool: tensor/expert parallelism, topology and per-device peaks must be recorded.
A consistent test across all three tiers
- Record the exact repository, revision, artifact and projector.
- Fix runtime, context allocation, cache precision, output target and concurrency.
- Measure peak accelerator and system memory during the real task.
- Capture prompt processing, time to first token and decode throughput separately.
- Evaluate output quality before treating a lower-bit fit as acceptable.
- Compare complete-system cost, power, cooling, support and return risk.
Use the repeatable workload test, inspect the fixed benchmark snapshot, or submit a controlled result. The database remains 43 metric observations from seven sources and does not establish a universal winner for these tiers.
Turn the map into a decision
Start with the VRAM calculator, browse the GPU comparison directory, then stage a browser-private watchlist. The current presets cover a 24B Q4/Q5 value path, Qwen3.8-27B 24GB/32GB community-GGUF candidates and gpt-oss-120b single-device candidates. Nothing is saved until the visitor explicitly chooses Save.
Browse the Weekly Digest archive for earlier decision maps. RSS is active; email signup and delivery are not currently active.
Sources and limits
- HighVRAM 24B buyer guide with linked official Mistral sources
- HighVRAM 27B buyer guide with linked Google/Qwen sources
- HighVRAM Qwen3.8-27B memory ladder with linked official and community artifact sources
- HighVRAM gpt-oss-120b guide with linked official OpenAI sources
- HighVRAM methodology
Published October 25, 2026. File sizes and memory guidance are tied to the sources linked in the underlying guides. HighVRAM has not independently benchmarked every configuration. Fit does not establish speed, output quality, reliability or value.