gpt-oss-120b Speeds: GB10 Boxes Within 15.3%, RTX PRO 6000 Peaks at 6.8x

These are the single-user figures independent labs have published for desk machines that run a 120-billion-parameter model. They are not one league table: the labs counted different things with different tools. Only one lab's numbers, on one harness and one token count, rank machines. On those terms, five GB10 boxes sit within 15.3% of each other, and an RTX PRO 6000 card put out 6.8 times the output tokens of a GB10 box at full load. The rest give a range.

The Common Test: gpt-oss-120b, Measured Four Ways

gpt-oss-120b is the shared yardstick because it fits almost everywhere. Its MXFP4 four-bit GGUF build is a 63.39 GB file, StorageReview reported, so it loads on any 128 GB unified-memory box and on a 96 GB RTX PRO 6000 card. We found four kinds of run.

The first is StorageReview's vLLM online serving sweep on GB10 boxes: equal input and output lengths, batch 1 to 64, total tokens counted. The second is StorageReview's newer vLLM harness for workstation GPUs, which reports output tokens (and total tokens separately) at up to 128 streams on 512-token prompts and 512-token outputs. The third is llama.cpp's server, which StorageReview used on the RTX PRO 6000 with 8,192-token prompts and 1,024-token outputs. The fourth is LM Studio, a desktop app run by one user, used by ServeTheHome, StorageReview and HotHardware on Strix Halo machines.

A tokens-per-second figure is only as good as its count. Some tests count only the words the machine writes, called output tokens. Others also count the prompt it reads, which goes through many times faster, so a total figure looks bigger. The second trap is the batch, the number of requests answered at once. At batch 1 one person gets the whole machine; at batch 64 the headline figure is shared among 64 requests. Two numbers compare only when the model, the software, the prompt length, the batch and the token count all match.

GB10 Boxes: 60.71 to 70 tok/s for One User

StorageReview published four GB10 reviews on March 3, 2026, all on the same vLLM sweep, counting prompt and output tokens. At batch 1, the ASUS Ascent GX10 ran 70 tok/s, the Acer Veriton GN100 69.65, the Dell Pro Max with GB10 68.57 and the Gigabyte AI TOP ATOM 60.71. The HP ZGX Nano, reviewed on April 24, ran 69.6. Top to bottom, that is a 15.3% spread at batch 1 (70 divided by 60.71, our arithmetic). StorageReview did not state the quantization for the sweep; gpt-oss-120b ships in MXFP4.

The batch-64 figures are in our Nov. 2 GB10 comparison, from HP's 722.9 tok/s down to Gigabyte's 649.62. That story did not compute the multiple. Every box multiplied its batch-1 figure by 9.71 to 10.70 times: ASUS 9.71 (680 over 70), Dell 9.78 (670.83 over 68.57), Acer 10.24 (713.18 over 69.65), HP 10.39 (722.9 over 69.6) and Gigabyte 10.70 (649.62 over 60.71). Split 64 ways, each request gets 10.15 to 11.30 tok/s of total throughput (649.62 and 722.9 over 64), so a shared box serves a team, but each member waits far longer than a lone user.

All five share NVIDIA's 128 GB of memory at 273 GB/s. The HP run came 52 days after the others, and StorageReview did not say whether newer software helped it. The Acer review and the ASUS review carry the thermals and storage that do separate them.

Strix Halo, RTX PRO 6000 and the Mac Studio Against Them

The one same-lab test of AMD's Strix Halo against GB10 is StorageReview's July 6 Ryzen AI Halo review. At batch 64, with 256 tokens in and 256 out, the Halo served 222 tok/s against 701 for the DGX Spark, a gap of 3.16 times (701 over 222). StorageReview did not state the quantization. The gap depends on the workload: with 8,192-token prompts and 1,024-token outputs the Spark reached 2,760 against the Halo's 314, about 8.8 times, and with 1,024 tokens in and 8,192 out it was 305 against 127, about 2.4 times, StorageReview reported. AMD lists 256 GB/s and NVIDIA 273 GB/s; a bandwidth difference of about 7% (273 over 256) cannot explain a 3.16x gap, which points to compute and software, as our Halo review argued.

For one user, the Strix Halo figures come from LM Studio. ServeTheHome measured about 45 tok/s on the Framework Desktop in LM Studio's default setup. StorageReview measured close to 40 on HP's Z2 Mini G1a with 96 GB assigned to the GPU. HotHardware reported about 24 on the same HP machine and app in a Sept. 24, 2025 review, with 64 GB the largest practical GPU-memory setting, as our Z2 Mini review reported. That is a 1.9 times range (45 over 24) for one chip in one app. Settings and software versions differ, so it does not rank the machines.

In StorageReview's Oct. 20 review of Dell's Pro Precision 9 T2 tower, llama.cpp served gpt-oss-120b at 182.7 output tok/s to one user on the 600W Workstation Edition and 160.6 on the 300W Max-Q, after 8,192-token prompts; the review did not state the quantization. With 64 concurrent requests, the 600W card moved 3,805 tok/s counting prompt and output tokens, but each request received 6.6 output tok/s, and output across all 64 came to 422.8 tok/s, StorageReview measured. That is 2.3 times the single-user 182.7 (422.8 over 182.7, our arithmetic). The 182.7 is not comparable with the GB10's 70: it counts only output tokens, on a different tool and with 8,192-token prompts.

The like-for-like comparison comes from StorageReview's Aug. 24 review of MSI's XpertStation WS300, which ran an RTX PRO 6000 Workstation Edition and a Spark-class GB10 unit, an Acer GN100, on one vLLM harness with 512 tokens in and 512 out. Counting output tokens at peak, over 1 to 128 streams, the card reached 3,384 tok/s against 500 for the GB10 box, 6.8 times as much (3,384 over 500, our arithmetic), while the GB300-based WS300 reached 9,294. NVIDIA rates the card at 96 GB and 1,792 GB/s. Our RTX PRO 6000 review has its heat issues, and owners have reported 12V-2x6 cable damage.

We found no gpt-oss-120b result from a recognized lab for Apple's Mac Studio with M5 Ultra, which Apple lists at 1.2 TB/s, more than four times a GB10's 273 GB/s. The nearest Apple figure is the M5 Max: tbreak reported 75.65 tok/s on gpt-oss-120b in Ollama on a 128 GB Mac Studio on Sept. 21, against 44.06 for an ASUS Ascent GX10 from the reviewer's earlier testing. Tom's Hardware's Nov. 10 M5 Ultra tests, which we compared with the Spark, used other models, with figures shown only in its charts. On Qwen 3.8 27B in llama.cpp the Mac finished an agent replay in 555.2 seconds against 1,020.2 for a Dell GB10. On GLM-5.3-Flash, with the TensorFold engine, an MLX 4-bit build and multi-token prediction, it wrote 94.0 tok/s at zero context.

Why the Numbers Differ, and Which Ones Compare

StorageReview's own runs show the trap. In July, a DGX Spark served 701 tok/s at batch 64 on gpt-oss-120b. In the Aug. 24 WS300 review, the same lab put a Spark-class GB10 unit at 500 output tok/s at peak. The 701 and the 500 differ mainly in what they count: the July figure matches the March sweep, which counts prompt and output tokens, while the August figure counts output only, with longer prompts, more streams, a different GB10 unit and newer software.

Numbers drift even within one harness. StorageReview logged the RTX PRO 6000 at a peak of 3,384 output tok/s in August and 3,064 at 128 streams on Nov. 3, when it tested the 600W Server Edition beside the DGX Station's 9,280. The August figure is 10.4% higher (3,384 over 3,064). The card edition and host changed in between.

What compares is narrow. The five GB10 boxes compare with each other on StorageReview's March and April total-token sweep. The Spark and the Halo compare at batch 64 in July. Within each of StorageReview's August and November reviews, the RTX PRO 6000 compares with the WS300 or the DGX Station on output tokens at up to 128 streams, and in August with a GB10 box. Single-user LM Studio figures on Strix Halo give a range, not a rank.

For engineers sizing a deployment, the shared-load figure matters more than the single-user one. On the one harness that ran both, the RTX PRO 6000 put out 6.8 times a GB10 box's output tokens at peak, so the card serves a shared team far better. A GB10 box's 9.71 to 10.70 times gain from batch 1 to 64 is a total-token multiple and does not compare with the card's 2.3 times output-token gain in llama.cpp. For investors, the spread inside the GB10 family is 11.3% at batch 64 (722.9 over 649.62) and 15.3% at batch 1, which leaves price, stock and support to decide sales. Before comparing any two figures, ask for the tool, the batch and the token count.

What to Watch

Apple's 512 GB Mac Studio configuration is due Dec. 3, as our review reported, and a gpt-oss-120b run on any M5 Ultra would fill the largest gap here. Framework listed Batch 3 of its 192 GB Desktop for December; independent tests will show what its 273 GB/s changes on Strix Halo. NVIDIA says DGX Station for Windows arrives in December 2026. MSI's RTX Spark mini PC, the EdgeMesa N, lists at $7,999 but was out of stock on Nov. 11, as we reported, and we found no published gpt-oss-120b run on it. On GB10, the kernel 7.0 freezes we reported on Nov. 9 were still open, and any fix could move the numbers. None of the March GB10 sweeps has been rerun on current software.