gpt-oss 20B is 21.5B total / 3.6B active, 131,072 tokens on the official card. Llama 3.3 70B is 70.6B, 128K tokens. Q6 / Q5 / Q4 file sizes come from the listed GGUFs on each /models page, or a calculator estimate when a listing is missing. Q3 stays off. Q4 fits gpt-oss 20B on one R9700 at 4K and spills on Llama 3.3 70B. Neither model was measured on the local desktop. Fit numbers are calculator estimates. llama.cpp first.
How we test
— Measured cells cite official amdai-bench/run@1 cards. Estimates stay labeled. Q3 is estimate-only. Q1/Q2 stay off.
Official card: 21B parameters with 3.6B active. Official config.json max_position_embeddings is 131,072. Alternating attention: 12 full-context layers, 12 at a 128-token window. Native MXFP4 MoE weights keep GGUF Q6/Q5/Q4/Q3 files close in size; the Q3 row is a labeled estimate.
Existing GPU sheets only: R9700, RX 7900 XTX, RX 7900 XT, RX 7800 XT. llama.cpp, f16 KV.
Totals are calculator estimates. Speed is measured only on R9700 35B cells that already exist;
everything else stays not measured yet. This table lists Q6 / Q5 / Q4. Q3 stays on the calculator as a labeled estimate.
No official local desktop cell for this model × card. Not measured yet.
12.6 GB
Fits in VRAM (estimate)
not measured yet
No official local desktop cell for this model × card. Not measured yet.
14.8 GB
Fits in VRAM (estimate)
not measured yet
No official local desktop cell for this model × card. Not measured yet.
56.4 GB
Does not fit this GPU
not measured yet
No official local desktop cell for this model × card. Not measured yet.
65.3 GB
Does not fit this GPU
not measured yet
No official local desktop cell for this model × card. Not measured yet.
94.8 GB
Does not fit this GPU
not measured yet
No official local desktop cell for this model × card. Not measured yet.
Q5
Q5_K_M / Q5_K_M
Estimate
11.6 GB
Fits in VRAM (estimate)
not measured yet
No official local desktop cell for this model × card. Not measured yet.
12.3 GB
Fits in VRAM (estimate)
not measured yet
No official local desktop cell for this model × card. Not measured yet.
14.5 GB
Fits in VRAM (estimate)
not measured yet
No official local desktop cell for this model × card. Not measured yet.
48.9 GB
Does not fit this GPU
not measured yet
No official local desktop cell for this model × card. Not measured yet.
57.8 GB
Does not fit this GPU
not measured yet
No official local desktop cell for this model × card. Not measured yet.
87.3 GB
Does not fit this GPU
not measured yet
No official local desktop cell for this model × card. Not measured yet.
Q4
Q4_K_M / Q4_K_M
Estimate
11.5 GB
Fits in VRAM (estimate)
not measured yet
No official local desktop cell for this model × card. Not measured yet.
12.2 GB
Fits in VRAM (estimate)
not measured yet
No official local desktop cell for this model × card. Not measured yet.
14.4 GB
Fits in VRAM (estimate)
not measured yet
No official local desktop cell for this model × card. Not measured yet.
41.9 GB
Does not fit this GPU
not measured yet
No official local desktop cell for this model × card. Not measured yet.
50.7 GB
Does not fit this GPU
not measured yet
No official local desktop cell for this model × card. Not measured yet.
80.2 GB
Does not fit this GPU
not measured yet
No official local desktop cell for this model × card. Not measured yet.
System RAM
When a shopper file does not fit, the spill is system RAM.
Shopper lines stay Minimum system RAM needed and
RAM needed at full context on one R9700 at the model’s shopper quant,
32K selected and the official-card full context. Estimate unless a measured R9700 card overlays the kit.
gpt-oss 20B
Q4_K_M at 4K and 32K fits the listed cards on the calculator.
Minimum system RAM needed
32 GB kit
Estimate · 24.6 GiB raw · 32K context
RAM needed at full context
32 GB kit
Estimate · 24.6 GiB raw · 8K full context
Estimate. Offloaded weights + KV in RAM + runtime + OS/desktop + headroom, rounded up to a 32 GB DIMM kit.
Breakdown
Offloaded weights ·
0 GiB
Separable embed in RAM ·
0 GiB
KV in RAM at selected ·
0 GiB
KV in RAM at full context ·
0 GiB
Runtime / backend ·
0.57 GiB
OS / desktop baseline ·
8.00 GiB
Headroom ·
16.0 GiB
Llama 3.3 70B
Q4_K_M at 4K or 32K does not fit every listed card on the calculator.
Minimum system RAM needed
48 GB kit
Estimate · 42.7 GiB raw · 32K context
RAM needed at full context
48 GB kit
Estimate · 35.1 GiB raw · 8K full context
Estimate. Offloaded weights + KV in RAM + runtime + OS/desktop + headroom, rounded up to a 48 GB DIMM kit.
Breakdown
Offloaded weights ·
7.60 GiB
Separable embed in RAM ·
0 GiB
KV in RAM at selected ·
10.0 GiB
KV in RAM at full context ·
2.50 GiB
Runtime / backend ·
1.14 GiB
OS / desktop baseline ·
8.00 GiB
Headroom ·
16.0 GiB
llama.cpp first
Commands and recipes already on the model pages. Cited-model shapes are not measured local desktop recipes.
House Flash keeps its existing Strata recipe plus the llama.cpp backend page.
# Shape only — not a measured local desktop recipe.
./build-vk/bin/llama-server \
-m ~/models/Llama-3.3-70B-Instruct-Q4_K_M.gguf \
-ngl 99 \
-c 8192 \
--cache-type-k f16 \
--cache-type-v f16 \
--host 127.0.0.1 --port 8080
Plain-English pick
Chat
gpt-oss 20B holds Q4 on one R9700 at 4K. Llama 3.3 70B needs RAM offload.
Coding agents
Keep llama.cpp first. The side that holds Q4 in VRAM (gpt-oss 20B) is the less painful coding-agent host. House Flash / agent pins stay R9700 / local desktop — they do not transfer. No invented agent speed.
Long context
Cited context is 131,072 tokens vs 128K tokens. KV grows with context. Q4 at 32K fits gpt-oss 20B and spills on Llama 3.3 70B on one R9700. Read /long-context for measured R9700 35B 128K–262K cards. We do not invent a 128K decode figure here.
Size and VRAM
gpt-oss 20B files: Q6_K 12.04 GB, Q5_K_M 11.72 GB, Q4_K_M 11.62 GB. Llama 3.3 70B files: Q6_K 57.89 GB, Q5_K_M 49.95 GB, Q4_K_M 42.52 GB. Fit on the four existing GPU sheets below. Empty speed cells stay not measured yet.
Stack Lab
Open in Stack Lab only where the handoff is already allowlisted. Doctor is always here.
House daily / Flash / agent profiles appear only when that house model is on the pair.
If the app is not installed, packages are not published yet. There is no public download URL. Read the
Stack Lab install notes
.
If Stack Lab is already running on this machine,
open the local app
.
Email only — no account. Double opt-in. About once a week, and only when run cards, guides, ROCm problems, or Stack Lab actually changed. Not a daily drip.
Issues may include one labeled Sponsored mention.
Slots and rates:
Sponsor.