LocalsOnlyevaluationsHow to read this
← Model evidence

LocalsOnly · measured2026-10-14

GPQA Diamond

Diamond · version 1.0 · 198 tasks. Compare only matched task sets and protocols.

GLM-5.3-Flash · EXL3 TR3 4bpw · 2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total · TensorFold

Needs two desk boxes: DGX Spark (serve head) + ASUS GX10 (tensor-parallel worker), 128GB each.

● Tested by LocalsOnly
Download weights · Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw @ 25a44fdb ↗Serve recipe · MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks-TensorFold@549fdc5 ↗
WALL-CLOCK TIME29.9 h
INPUT TOKENS13,174,753

Recorded effort for this run. Wall time and token counts are not a dollar cost; hardware power and utilization also matter.

Reported score87.4%173 of 198 tasks passed
Throughput—

Tokens per second · see measurement basis

Concurrency1

As recorded on this receipt

0 infrastructure outcomes are recorded. Failure counts may overlap with task outcomes.
Read before comparing
  • Engine A/B: this row and the vLLM GPQA Diamond row (2026-09-29, mean 0.8586, 170/198) use the same Mia-AiLab EXL3 TR3 4bpw weights file @25a44fdb and an identical Harbor config (only job_name differs): terminus-2, thinking on, max_turns 50, k=1 n=1, full 198, 100,000-token cap, same two boxes. The serving stack changed: TensorFold v1.10 instead of vLLM.
  • Not a pure engine comparison: TensorFold serves the dense/shared weights at 4-bit, while the vLLM run kept them at BF16. TensorFold's server-default sampling was not separately verified to match vLLM's.
  • Mostly the same tasks: 166 tasks pass on both; 7 pass only on TensorFold and 4 only on vLLM. With k=1 n=1 the 3-task gap (173 vs 170) is within run-to-run noise (standard error about 2.4 points).
  • Timeout errors count as scored: 17 AgentTimeoutErrors (3600 s each); 16 scored 0 and 1 had already written a correct answer and scored 1. 6 of them (tasks 25, 48, 71, 94, 102, 121) hung with no model output. The vLLM run had 23 (20 scored 0, 3 scored 1).
  • Wall time: 29h53m here vs 38h32m for the vLLM run. Both were clean serial runs (0 retries, no resume), but the gap is not a speed measurement: timeout waits dominate and TensorFold generated more tokens (3.40M vs 2.23M output).
  • Speed: a separate post-run single-stream probe on 10/14 (same speed.py prompts and settings as the 10/9 speed test, thinking off, 512-token generations, 3 runs) recorded TensorFold 49.7/49.7/49.8 tok/s, in line with the 10/9 TensorFold 50.4. The vLLM side was not re-probed; its 10/9 figure was 24.0 tok/s (mean of 23.2, 25.8, 23.1). Not measured during this GPQA run.
  • Weights are Mia-AiLab EXL3 TR3 4bpw (community quant) with a DFlash2 speculator, fp8 KV cache; not Z.ai's official weights. Harbor terminus-2 agentic harness (writes /app/answer.txt), not a one-shot multiple-choice prompt.
  • Needs two desk boxes: serve head DGX Spark spark-5c2f (:8888) + worker ASUS GX10 gx10-e789 (2× GB10, 256GB total). Not a single 128GB desk box.

Reported category outcomes

Each category keeps its own recorded denominator. Some categories exclude unscored trials. Mean reward, when present, is partial credit.

gpqa-diamond
87.4%
173 / 19887.4% passed · mean reward 87.4%
Run configuration & provenance +
Context
100,000
Temperature
Not recorded
Wall time
1793.1 minutes
Revision
25a44fdbf16862a46b7cc9921142c6c81350af2f
Reasoning
thinking on · server-default reasoning effort (not passed by harness)
Agent version
Not reported
Output token limit
Not reported
Dataset revision
harbor gpqa-diamond@1.0 (198 tasks) · harbor-datasets 1983ac5c4d43f43cb7a9af9f89c54d09025589ec
Harness host
mac
Model host
2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total
Input tokens
13,174,753
Output tokens
3,404,783
Cached input tokens
0
Agent turn / step limit
50
Concurrency
1
Weight file size
175.72 GB

Throughput measurement: Not recorded as a run throughput. Job wall time includes 17 x 3600 s timeout waits (6 of them hung with no model output), so wall-clock throughput is not a speed measurement. See speed_probe for a separate post-run single-stream probe.

Harbor terminus-2 on Kris Mac mini (docker) -> TensorFold v1.10 OpenAI-compat [local endpoint] (GLM-5.3-Flash-EXL3 TR3 4bpw, dense/shared 4-bit, fp8 KV, DFlash2 bf582e4e, context 100000, image v0.6.0-a1897d591f70, recipe 549fdc5; DGX Spark spark-5c2f head + ASUS GX10 gx10-e789 worker over ConnectX-7; thinking on, max_turns=50, k=1 n=1, server-default sampling, 3600 s task timeout). Job gpqa-mac-dgx-glm53flash-tf110-20261013a. config.json identical to the vLLM job gpqa-mac-dgx-glm53flash-20260928a except job_name (diffed). Dataset gpqa-diamond@1.0.

Run ID: measured-2x-gb10-glm53flash-exl3-4bpw-tensorfold-gpqa-diamond-1.0-2026-10-14
Tasks

Task explorer

Passed tasks first, then other outcomes grouped by recorded issue. Use the filters to inspect a particular outcome.

198 task records
173 Passed25 Not passed

All 198 recorded tasks, before filters or pagination. 25 of 198 matching tasks shown below. Recorded issues can also occur on scored tasks.

PassedTask ID 0+

Task ID: 0
Recorded name: 0

Reward: 1

Tokens in / out: 2,282 / 1,079

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 1+

Task ID: 1
Recorded name: 1

Reward: 1

Tokens in / out: 2,354 / 3,860

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 2+

Task ID: 2
Recorded name: 2

Reward: 1

Tokens in / out: 2,370 / 791

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 3+

Task ID: 3
Recorded name: 3

Reward: 1

Tokens in / out: 2,259 / 1,201

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 4+

Task ID: 4
Recorded name: 4

Reward: 1

Tokens in / out: 8,321 / 2,046

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 5+

Task ID: 5
Recorded name: 5

Reward: 1

Tokens in / out: 2,478 / 2,709

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 6+

Task ID: 6
Recorded name: 6

Reward: 1

Tokens in / out: 2,634 / 3,298

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 7+

Task ID: 7
Recorded name: 7

Reward: 1

Tokens in / out: 2,933 / 17,456

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 9+

Task ID: 9
Recorded name: 9

Reward: 1

Tokens in / out: 2,307 / 1,596

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 10+

Task ID: 10
Recorded name: 10

Reward: 1

Tokens in / out: 1,191,168 / 67,378

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 11+

Task ID: 11
Recorded name: 11

Reward: 1

Tokens in / out: 4,300 / 1,522

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 14+

Task ID: 14
Recorded name: 14

Reward: 1

Tokens in / out: 2,708 / 1,064

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 16+

Task ID: 16
Recorded name: 16

Reward: 1

Tokens in / out: 2,476 / 1,498

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 18+

Task ID: 18
Recorded name: 18

Reward: 1

Tokens in / out: 3,176 / 20,373

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 19+

Task ID: 19
Recorded name: 19

Reward: 1

Tokens in / out: 2,360 / 915

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 20+

Task ID: 20
Recorded name: 20

Reward: 1

Tokens in / out: 2,640 / 11,862

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 21+

Task ID: 21
Recorded name: 21

Reward: 1

Tokens in / out: 2,238 / 7,747

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 22+

Task ID: 22
Recorded name: 22

Reward: 1

Tokens in / out: 2,563 / 13,036

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 23+

Task ID: 23
Recorded name: 23

Reward: 1

Tokens in / out: 2,492 / 27,551

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 24+

Task ID: 24
Recorded name: 24

Reward: 1

Tokens in / out: 2,742 / 4,817

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 26+

Task ID: 26
Recorded name: 26

Reward: 1

Tokens in / out: 2,350 / 1,078

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 27+

Task ID: 27
Recorded name: 27

Reward: 1

Tokens in / out: 76,459 / 74,057

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 28+

Task ID: 28
Recorded name: 28

Reward: 1

Tokens in / out: 7,676 / 4,671

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 30+

Task ID: 30
Recorded name: 30

Reward: 1

Tokens in / out: 2,216 / 26,346

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
PassedTask ID 31+

Task ID: 31
Recorded name: 31

Reward: 1

Tokens in / out: 2,665 / 5,964

Recorded issue: No failure explanation in this task record.

The receipt contains outcome metadata. Full prompts and execution transcripts are not included in this view.
Showing 25 of 198 tasks