DVIDIA
DVIDIA × OWterminal

Open weights.
A place to start.

Find a model. Understand the hardware. Make it useful.

DVIDIA is where skills become useful work. Our sister site, OWterminal, helps you research the models and machines behind them.

Find a setup Explore hardware

01 / Explore the possibilities

What do you want to run?

Full comparison desk

16 community-reported setups from OWterminal · Reports 2026-09-16–2026-09-17. Saved 2026-09-21. Speeds depend on the exact build and test settings; DVIDIA has not reproduced these reports.

16 reported setups

Every result has a source.
How rankings & scores work

Rank orders the reported output speeds in these results. Speed index = reported tokens/sec ÷ fastest visible reported tokens/sec × 100. The fastest visible setup scores 100; equal speeds share a rank. Filtering recalculates both. Reports use different models, prompts, context lengths and builds, so compare the source conditions before deciding.

Qwen3.5 2B

Q4_K_M GGUF

#1
RTX 5090 32GBClaimed
Reported tokens/sec
351
Speed index
100 / 100
Runtime
Not reported
Reported peak memory
Not reported
Details for Qwen3.5 2B on RTX 5090 32GB

@fillagrew · 2026-09-17

351 tokens/sec on Qwen3.5 2B at Q4_K_M
Build
Q4_K_M GGUF

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @fillagrew
Inspect setup : Qwen3.5 2B on RTX 5090 32GB, on OWterminal

Gemma-4 26B-A4B-it

Q4_K_M GGUFabliterated

#2
RTX 5090 32GBClaimed
Reported tokens/sec
173
Speed index
49.3 / 100
Runtime
Not reported
Reported peak memory
Not reported
Details for Gemma-4 26B-A4B-it on RTX 5090 32GB

@fillagrew · 2026-09-17

173 tok/s on Gemma-4 26B-A4B-it-abliterated at Q4_K_M
Build
Q4_K_M GGUF

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @fillagrew
Inspect setup : Gemma-4 26B-A4B-it on RTX 5090 32GB, on OWterminal

Qwen3.8-Flash-Next

Build not specified

#3
M5 MaxClaimed
Reported tokens/sec
126.5
Speed index
36 / 100

peak · 61 @100k · 50 @200k

Runtime
MTPLX V2.11.3
Reported peak memory
Not reported
Details for Qwen3.8-Flash-Next on M5 Max

@Youssofal_ · 2026-09-17

Peak Speed: 126.5 TPS.
Build
Build not specified
Reported test context
200,000 tokens

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @Youssofal_
Inspect setup : Qwen3.8-Flash-Next on M5 Max, on OWterminal

Qwen3.6-35B-A3B

NVFP4abliterated

#4
2× DGX SparkClaimed
Reported tokens/sec
94.4
Speed index
26.9 / 100
Runtime
Not reported
Reported peak memory
Not reported
Details for Qwen3.6-35B-A3B on 2× DGX Spark

@bonellisystems · 2026-09-17

On our 2× DGX Sparks the Qwen3.6-35B-A3B abliterated NVFP4 gauntlet still reads 94.4 decode tok/s and 181.5 prefill with TTFT 143ms.
Build
NVFP4

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @bonellisystems
Inspect setup : Qwen3.6-35B-A3B on 2× DGX Spark, on OWterminal

Qwen3.6-35B-A3B

4-bit MLX

#5
M4 ProClaimed
Reported tokens/sec
85.5
Speed index
24.4 / 100
Runtime
rapid-mlx
Reported peak memory
21 GB RAM
Details for Qwen3.6-35B-A3B on M4 Pro

@rapidmlx · 2026-09-16

qwen3.6-35b-a3b → 85.5 tok/s
Build
4-bit MLX

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @rapidmlxBase model card
Inspect setup : Qwen3.6-35B-A3B on M4 Pro, on OWterminal

Qwen3.5-4B

4-bit MLX

#6
M4 ProClaimed
Reported tokens/sec
82.8
Speed index
23.6 / 100
Runtime
rapid-mlx
Reported peak memory
8 GB RAM
Details for Qwen3.5-4B on M4 Pro

@rapidmlx · 2026-09-16

qwen3.5-4b → 82.8
Build
4-bit MLX

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @rapidmlx
Inspect setup : Qwen3.5-4B on M4 Pro, on OWterminal

Empero Qwen3.8-35B-A3B

Q4_K_M GGUF

#7
RTX 3060 12GBClaimed
Reported tokens/sec
≈ 50
Speed index
14.2 / 100
Runtime
llama.cpp
Reported peak memory
Not reported
Details for Empero Qwen3.8-35B-A3B on RTX 3060 12GB

@Oluwaphilemon1 · 2026-09-17

Empero’s Qwen3.8-35B-A3B is running on: RTX 3060 12GB 16GB system RAM 262K context ~550 tok/s prefill ~50 tok/s decode
Build
Q4_K_M GGUF
Machine system RAM
16 GB
Claimed context
262,144 tokens

Qwen3.8 distilled into Qwen3.6-35B-A3B. Not official Qwen3.8-27B.

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @Oluwaphilemon1
Inspect setup : Empero Qwen3.8-35B-A3B on RTX 3060 12GB, on OWterminal

Qwen3.5-9B

4-bit MLX

#8
M4 ProClaimed
Reported tokens/sec
49.3
Speed index
14 / 100
Runtime
rapid-mlx
Reported peak memory
Not reported
Details for Qwen3.5-9B on M4 Pro

@rapidmlx · 2026-09-16

qwen3.5-9b → 49.3
Build
4-bit MLX

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @rapidmlx
Inspect setup : Qwen3.5-9B on M4 Pro, on OWterminal

Qwen3-8B

4-bit MLX

#9
M4 ProClaimed
Reported tokens/sec
48.3
Speed index
13.8 / 100
Runtime
rapid-mlx
Reported peak memory
Not reported
Details for Qwen3-8B on M4 Pro

@rapidmlx · 2026-09-16

qwen3-8b → 48.3
Build
4-bit MLX

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @rapidmlx
Inspect setup : Qwen3-8B on M4 Pro, on OWterminal

Qwen3.8-Flash-Next

FP8

#10
2× DGX SparkClaimed
Reported tokens/sec
≈ 45
Speed index
12.8 / 100

Sustained with MTP, as reported

Runtime
Not reported
Reported peak memory
Not reported
Details for Qwen3.8-Flash-Next on 2× DGX Spark

@Oluwaphilemon1 · 2026-09-17

~45 tok/s sustained with MTP
Build
FP8
Claimed context
256,000 tokens

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @Oluwaphilemon1
Inspect setup : Qwen3.8-Flash-Next on 2× DGX Spark, on OWterminal

Qwen3.8-27B

UD-Q4_K_XL GGUF

#11
RTX 4090 24GBClaimed
Reported tokens/sec
40.7
Speed index
11.6 / 100

60.1 MTP @130k

Runtime
llama.cpp
Reported peak memory
23.7 GB VRAM
Details for Qwen3.8-27B on RTX 4090 24GB

@Oluwaphilemon1 · 2026-09-17

40.7 tok/s decode
Build
UD-Q4_K_XL GGUF
Reported test context
142,000 tokens
Claimed context
260,000 tokens

Not Empero 35B-A3B.

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @Oluwaphilemon1Base model card
Inspect setup : Qwen3.8-27B on RTX 4090 24GB, on OWterminal

Ornith

Build not specified

#12
GTX 1660 SUPER 6GBClaimed
Reported tokens/sec
≈ 40
Speed index
11.4 / 100

40–46 range

Runtime
Not reported
Reported peak memory
Not reported
Details for Ornith on GTX 1660 SUPER 6GB

@mine_craft_bui · 2026-09-17

ended up making a .sh script to get ~40–46 tok/s on a GTX 1660 SUPER
Build
Build not specified
Machine system RAM
40 GB

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @mine_craft_buiPublisher models
Inspect setup : Ornith on GTX 1660 SUPER 6GB, on OWterminal

Qwen3.8-27B

GSQ-RCO IQ3_XXS-mtp

#13
RTX 5060 Ti 16GBClaimed
Reported tokens/sec
23
Speed index
6.6 / 100
Runtime
llama.cpp
Reported peak memory
Not reported
Details for Qwen3.8-27B on RTX 5060 Ti 16GB

@fntAInhead · 2026-09-17

GSQ-RCO IQ3_XXS-mtp 23 tok/s
Build
GSQ-RCO IQ3_XXS-mtp
Claimed context
100,000 tokens

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @fntAInheadBase model card
Inspect setup : Qwen3.8-27B on RTX 5060 Ti 16GB, on OWterminal

Qwen3.8-27B

UD-Q2_K_XL

#14
RTX 5060 Ti 16GBClaimed
Reported tokens/sec
19
Speed index
5.4 / 100
Runtime
llama.cpp
Reported peak memory
Not reported
Details for Qwen3.8-27B on RTX 5060 Ti 16GB

@fntAInhead · 2026-09-17

UD-Q2_K_XL 19 tok/s
Build
UD-Q2_K_XL
Claimed context
128,000 tokens

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @fntAInheadBase model card
Inspect setup : Qwen3.8-27B on RTX 5060 Ti 16GB, on OWterminal

Nex-N2.5-mini

Q4_K_M GGUF

#15
RTX 5060 Ti 16GBClaimed
Reported tokens/sec
14
Speed index
4 / 100
Runtime
llama.cpp
Reported peak memory
Not reported
Details for Nex-N2.5-mini on RTX 5060 Ti 16GB

@fntAInhead · 2026-09-17

Nex Q4_K_M 14 tok/s
Build
Q4_K_M GGUF
Claimed context
262,000 tokens

35B-A3B post-train, not a 27B.

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @fntAInhead
Inspect setup : Nex-N2.5-mini on RTX 5060 Ti 16GB, on OWterminal

Qwen3.8-27B TurboFCFusion

IQ2_M GGUFuncensored

#16
RTX 5060 Ti 16GBClaimed
Reported tokens/sec
10
Speed index
2.8 / 100
Runtime
llama.cpp
Reported peak memory
Not reported
Details for Qwen3.8-27B TurboFCFusion on RTX 5060 Ti 16GB

@fntAInhead · 2026-09-17

TurboFC IQ2_M 10 tok/s
Build
IQ2_M GGUF
Claimed context
100,000 tokens

Speed index compares reported generation speed in the current results; 100 is the fastest. Builds and test conditions differ.

Exact build revision and artifact license are not linked in this report.

Source report by @fntAInhead
Inspect setup : Qwen3.8-27B TurboFCFusion on RTX 5060 Ti 16GB, on OWterminal

Hardware selection finds reports for that machine; it does not establish that a model fits your exact configuration. Missing memory or runtime details remain visible in each result.

02 / Put it to work

Choose your next step.

Choose hardware

Research before you buy.

OWterminal brings hardware comparisons and sourced market listings together, so you can inspect the evidence and the seller’s current terms.

Compare hardware Browse the hardware market
Build with DVIDIA

Give the model a useful task.

Explore an open skill or record a demonstration. Check its artifact type and runtime requirements before pairing it with a model.

Explore skills Record a demonstration

Coming next / The provider network

Bring your own capacity.

Bring an open-weight model endpoint through the DVIDIA CLI. Local provider preparation and validation are implemented; public distribution and verified network enrollment are next.

The planned service links request usage to a verified provider address, then shares collected customer revenue. Network-token incentives will be tracked separately; rates, eligibility and rewards are not live.

  • Open model licenses stay visible
  • Pay for delivered service
  • Contributors share in collected revenue