Rev. 2026.10

Rankings

Pick your memory class. Speed and quality sit side by side and are never folded into one magic score, so you can see the trade-off yourself.

Estimates included · filter for measured runs

Measured runs only

4 matching models. Includes formula estimates. This ordering is not a measured leaderboard.

16 GB class, on GeForce RTX 5060 Ti 16 GB

Default quant, 8k context, single stream. Fastest first.

#	Model	Quant	Memory	Verdict	Decode, tok/s	Quality score
1	gpt-oss-20b	MXFP4	12.9 GB	Fits	117.4Estimate	1,287Sourced
2	Llama 3.1 8B Instruct	Q4_K_M	6.6 GB	Fits	44.8Estimate	1,187Sourced
3	Qwen3 8B	Q4_K_M	6.8 GB	Fits	43.1Estimate	Not yet measured
4	Qwen3 14B	Q4_K_M	10.9 GB	Fits	26.0Estimate	Not yet measured

Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.
- Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
- Sourced Taken from a cited source; not yet independently checked.
- Estimate Calculated with the published formulas; not a measurement.
- Sample Placeholder data for layout; not a real figure.
- Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.
Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.
Evidence labels · Freshness policy · Calculation methodology

Measured runs only

8 matching models. Includes formula estimates. This ordering is not a measured leaderboard.

24 GB class, on GeForce RTX 3090 24 GB (used)

Default quant, 8k context, single stream. Fastest first.

#	Model	Quant	Memory	Verdict	Decode, tok/s	Quality score
1	gpt-oss-20b	MXFP4	12.9 GB	Fits	245.3Estimate	1,287Sourced
2	Qwen3 30B-A3B	Q4_K_M	19.9 GB	Fits	199.8Estimate	1,317Sourced
3	Llama 3.1 8B Instruct	Q4_K_M	6.6 GB	Fits	93.7Estimate	1,187Sourced
4	Qwen3 8B	Q4_K_M	6.8 GB	Fits	90.0Estimate	Not yet measured
5	Qwen3 14B	Q4_K_M	10.9 GB	Fits	54.3Estimate	Not yet measured
6	Mistral Small 3.2 24B	Q4_K_M	16.2 GB	Fits	35.8Estimate	1,338Sourced
7	Gemma 3 27B IT	Q4_K_M	18.2 GB	Fits	31.8Estimate	1,358Sourced
8	Qwen3 32B	Q4_K_M	22.5 GB	Tight	25.6Estimate	1,340Sourced

Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.
- Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
- Sourced Taken from a cited source; not yet independently checked.
- Estimate Calculated with the published formulas; not a measurement.
- Sample Placeholder data for layout; not a real figure.
- Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.
Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.
Evidence labels · Freshness policy · Calculation methodology

Measured runs only

8 matching models. Includes formula estimates. This ordering is not a measured leaderboard.

32 GB class, on GeForce RTX 5090 32 GB

Default quant, 8k context, single stream. Fastest first.

#	Model	Quant	Memory	Verdict	Decode, tok/s	Quality score
1	gpt-oss-20b	MXFP4	12.9 GB	Fits	469.6Estimate	1,287Sourced
2	Qwen3 30B-A3B	Q4_K_M	19.9 GB	Fits	382.4Estimate	1,317Sourced
3	Llama 3.1 8B Instruct	Q4_K_M	6.6 GB	Fits	179.4Estimate	1,187Sourced
4	Qwen3 8B	Q4_K_M	6.8 GB	Fits	172.4Estimate	Not yet measured
5	Qwen3 14B	Q4_K_M	10.9 GB	Fits	104.0Estimate	Not yet measured
6	Mistral Small 3.2 24B	Q4_K_M	16.2 GB	Fits	68.6Estimate	1,338Sourced
7	Gemma 3 27B IT	Q4_K_M	18.2 GB	Fits	60.9Estimate	1,358Sourced
8	Qwen3 32B	Q4_K_M	22.5 GB	Fits	49.1Estimate	1,340Sourced

Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.
- Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
- Sourced Taken from a cited source; not yet independently checked.
- Estimate Calculated with the published formulas; not a measurement.
- Sample Placeholder data for layout; not a real figure.
- Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.
Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.
Evidence labels · Freshness policy · Calculation methodology

Measured runs only

9 matching models. Includes formula estimates. This ordering is not a measured leaderboard.

48 GB class, on 2× RTX 3090 (used) 48 GB

Default quant, 8k context, single stream. Fastest first.

#	Model	Quant	Memory	Verdict	Decode, tok/s	Quality score
1	gpt-oss-20b	MXFP4	12.9 GB	Fits	245.3Estimate	1,287Sourced
2	Qwen3 30B-A3B	Q4_K_M	19.9 GB	Fits	199.8Estimate	1,317Sourced
3	Llama 3.1 8B Instruct	Q4_K_M	6.6 GB	Fits	93.7Estimate	1,187Sourced
4	Qwen3 8B	Q4_K_M	6.8 GB	Fits	90.0Estimate	Not yet measured
5	Qwen3 14B	Q4_K_M	10.9 GB	Fits	54.3Estimate	Not yet measured
6	Mistral Small 3.2 24B	Q4_K_M	16.2 GB	Fits	35.8Estimate	1,338Sourced
7	Gemma 3 27B IT	Q4_K_M	18.2 GB	Fits	31.8Estimate	1,358Sourced
8	Qwen3 32B	Q4_K_M	22.5 GB	Fits	25.6Estimate	1,340Sourced
9	Llama 3.3 70B Instruct	Q4_K_M	45.8 GB	Tight	12.4Estimate	1,274Sourced

Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.
- Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
- Sourced Taken from a cited source; not yet independently checked.
- Estimate Calculated with the published formulas; not a measurement.
- Sample Placeholder data for layout; not a real figure.
- Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.
Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.
Evidence labels · Freshness policy · Calculation methodology

Measured runs only

10 matching models. Includes formula estimates. This ordering is not a measured leaderboard.

128 GB unified class, on Ryzen AI Max+ 395 mini PC 128 GB

Default quant, 8k context, single stream. Fastest first.

#	Model	Quant	Memory	Verdict	Decode, tok/s	Quality score
1	gpt-oss-20b	MXFP4	12.9 GB	Fits	50.3Estimate	1,287Sourced
2	Qwen3 30B-A3B	Q4_K_M	19.9 GB	Fits	41.0Estimate	1,317Sourced
3	gpt-oss-120b	MXFP4	64.3 GB	Fits	37.5Estimate	1,365Sourced
4	Llama 3.1 8B Instruct	Q4_K_M	6.6 GB	Fits	19.2Estimate	1,187Sourced
5	Qwen3 8B	Q4_K_M	6.8 GB	Fits	18.5Estimate	Not yet measured
6	Qwen3 14B	Q4_K_M	10.9 GB	Fits	11.1Estimate	Not yet measured
7	Mistral Small 3.2 24B	Q4_K_M	16.2 GB	Fits	7.4Estimate	1,338Sourced
8	Gemma 3 27B IT	Q4_K_M	18.2 GB	Fits	6.5Estimate	1,358Sourced
9	Qwen3 32B	Q4_K_M	22.5 GB	Fits	5.3Estimate	1,340Sourced
10	Llama 3.3 70B Instruct	Q4_K_M	45.8 GB	Fits	2.5Estimate	1,274Sourced

Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.
- Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
- Sourced Taken from a cited source; not yet independently checked.
- Estimate Calculated with the published formulas; not a measurement.
- Sample Placeholder data for layout; not a real figure.
- Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.
Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.
Evidence labels · Freshness policy · Calculation methodology

Category rankings

One cited public benchmark per category; models without a comparable score are listed as not yet measured.

Each class uses one reference device. Cited benchmark runs only replace estimates when model, device, quant and recorded 8k context match. Sample speed figures are excluded. Unavailable quality scores are not filled with placeholders. Speed and quality are independent; vendor and community results may use different runtime settings. Special-purpose models have separate rankings by a cited public benchmark: start at Models.

Machine-readable: /api/v1/rankings/24gb.json and the other classes.