LocalsOnlyevaluationsHow to read this
← Model evidence

LocalsOnly · measured2026-10-28

Terminal-Bench 2.1

Version 2.1 · 89 tasks. Compare only matched task sets and protocols.

GLM-5.3-Flash · NVFP4/MXFP8 CSF-QAD · 2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total · Flare

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

● Tested by LocalsOnly
Download weights · local-inference-lab/GLM-5.3-Flash-NVFP4-MXFP8-CSF-QAD @ dec48abd ↗
WALL-CLOCK TIME44.6 h
INPUT TOKENSNot recorded

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

Reported score69.7%62 of 89 tasks passed
Throughput—

Tokens per second · see measurement basis

Concurrency1

As recorded on this receipt

3 infrastructure outcomes are recorded. Failure counts may overlap with task outcomes.
Read before comparing
  • Different engine, quant and context: NVIDIA Flare (commit 47581cc, flare 0.1.0) serving the local-inference-lab NVFP4/MXFP8 CSF-QAD checkpoint @dec48abd at a 1,048,576-token advertised context, vs EXL3 TR3 4bpw at a 100,000-token cap on the vLLM and TensorFold rows. Same suite and harness, so the rows sit together, but deltas are not engine-only.
  • Thinking forced ON: enable_thinking=false was sent (same as prior runs) but Flare ignores it; reasoning_content is present in all 1,551 agent steps across the 86 trajectories. The vLLM and TensorFold rows ran thinking off. This is the biggest recipe difference.
  • Sampling: server default temperature 1.0 / top_p 0.95; Flare clamps temperature to 0.95-1.0. Prior servers used their own defaults.
  • Single RoCE link between the boxes (B12X_ROCE_HCA=rocep1s0f1 pinned in the flare-agent drop-in) because the second link had transport retry errors.
  • CUDA 13.2 forward-compat libcuda (~/flare/cuda-compat via LD_LIBRARY_PATH) on host driver 580.173.02 (CUDA 13.0).
  • 3 harness errors (RuntimeError, no verifier reward) scored 0 and kept in the 89 denominator: overfull-hbox, qemu-alpine-ssh, qemu-startup. Excluding them gives 62/86 = 72.1% (secondary figure only, not the score).
  • Harbor flags 11 errored trials: the 3 RuntimeErrors plus 8 AgentTimeoutErrors (4x agent timeout). 7 timeouts verified at reward 0; torch-pipeline-parallelism timed out but verified at reward 1.
  • Endpoint reached through an SSH tunnel (Mac [local address]:18003 -> spark-5c2f [local address]:8003). Wall time includes tunnel and harness overhead; not a speed figure.
  • Not comparable to the official Terminal-Bench 2.1 card (Terminus, 256k, thinking on, no turn cap): this run keeps the 40-turn cap to mirror the prior local runs.
Run configuration & provenance +
Context
Not recorded
Temperature
1
Wall time
2678.8 minutes
Revision
dec48abd33efa73c3bb7c95b74eee10cad34f9be
Reasoning
on (forced by Flare)
Agent version
Not reported
Output token limit
Not reported
Dataset revision
sha256:7d7bdc1cbedad549fc1140404bd4dc45e5fd0ea7c4186773687d177ad3a0699a
Harness host
Not reported
Model host
2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total
Input tokens
Not reported
Output tokens
Not reported
Cached input tokens
Not reported
Agent turn / step limit
40
Concurrency
1
Weight file size
177.75 GB

Throughput measurement: Measurement method not separately reported. This should not be treated as standardized decode-only speed.

Harbor terminus-2 on Kris Mac mini (docker, native images) -> SSH tunnel [local address]:18003 -> NVIDIA Flare OpenAI-compat on spark-5c2f [local address]:8003 (GLM-5.3-Flash NVFP4/MXFP8 CSF-QAD @dec48abd, TP2, MTP draft tokens 3, flare-prep image 47581cc; DGX Spark spark-5c2f head + ASUS GX10 gx10-e789 worker over one RoCE link)

Run ID: measured-2x-gb10-glm53flash-nvfp4-flare-terminal-bench-2.1-2026-10-28
Tasks

Task explorer

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

89 task records
62 Passed24 Not passed3 Unscored · recorded issue

All 89 recorded tasks, before filters or pagination. 25 of 89 matching tasks shown below. Recorded issues can also occur on scored tasks.

PassedBn fit modify+

Task ID: bn-fit-modify
Recorded name: bn-fit-modify

Reward: 1

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.
PassedBreak filter js from html+

Task ID: break-filter-js-from-html
Recorded name: break-filter-js-from-html

Reward: 1

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.
PassedBuild cython ext+

Task ID: build-cython-ext
Recorded name: build-cython-ext

Reward: 1

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.
PassedBuild pmars+

Task ID: build-pmars
Recorded name: build-pmars

Reward: 1

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.
PassedBuild pov ray+

Task ID: build-pov-ray
Recorded name: build-pov-ray

Reward: 1

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.
PassedCaffe cifar 10+

Task ID: caffe-cifar-10
Recorded name: caffe-cifar-10

Reward: 1

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.
PassedCancel async tasks+

Task ID: cancel-async-tasks
Recorded name: cancel-async-tasks

Reward: 1

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.
PassedCircuit fibsqrt+

Task ID: circuit-fibsqrt
Recorded name: circuit-fibsqrt

Reward: 1

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.
PassedCobol modernization+

Task ID: cobol-modernization
Recorded name: cobol-modernization

Reward: 1

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.
PassedCode from image+

Task ID: code-from-image
Recorded name: code-from-image

Reward: 1

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.
PassedConfigure git webserver+

Task ID: configure-git-webserver
Recorded name: configure-git-webserver

Reward: 1

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.
PassedConstraints scheduling+

Task ID: constraints-scheduling
Recorded name: constraints-scheduling

Reward: 1

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.
PassedCount dataset tokens+

Task ID: count-dataset-tokens
Recorded name: count-dataset-tokens

Reward: 1

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.
PassedCrack 7z hash+

Task ID: crack-7z-hash
Recorded name: crack-7z-hash

Reward: 1

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.
PassedCustom memory heap crash+

Task ID: custom-memory-heap-crash
Recorded name: custom-memory-heap-crash

Reward: 1

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.
PassedDb wal recovery+

Task ID: db-wal-recovery
Recorded name: db-wal-recovery

Reward: 1

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.
PassedDistribution search+

Task ID: distribution-search
Recorded name: distribution-search

Reward: 1

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.
PassedDna insert+

Task ID: dna-insert
Recorded name: dna-insert

Reward: 1

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.
PassedExtract elf+

Task ID: extract-elf
Recorded name: extract-elf

Reward: 1

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.
PassedFeal differential cryptanalysis+

Task ID: feal-differential-cryptanalysis
Recorded name: feal-differential-cryptanalysis

Reward: 1

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.
PassedFeal linear cryptanalysis+

Task ID: feal-linear-cryptanalysis
Recorded name: feal-linear-cryptanalysis

Reward: 1

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.
PassedFix code vulnerability+

Task ID: fix-code-vulnerability
Recorded name: fix-code-vulnerability

Reward: 1

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.
PassedFix git+

Task ID: fix-git
Recorded name: fix-git

Reward: 1

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.
PassedFix ocaml gc+

Task ID: fix-ocaml-gc
Recorded name: fix-ocaml-gc

Reward: 1

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.
PassedGit leak recovery+

Task ID: git-leak-recovery
Recorded name: git-leak-recovery

Reward: 1

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 89 tasks