NoeFlandre/benchmark-llms-landuse-relevance

Land-use relevance benchmark

v3-multilingual · 85 languages x 300 items/language · 25,500 items · binary yes/no labels.

Code

Package version recorded in run metadata: 0.2.0 (some runs lack version metadata).

Task and prompt

Does a sentence describe a place's land or environment in ways visible to satellites?

English prompt · greedy decoding · seed 0 · max_new_tokens=4096 · bfloat16 · batch varies by model. unsloth/Qwen3.8-27B-GGUF@UD-IQ2_XXS runs the UD-IQ2_XXS GGUF quant through llama.cpp (same prompt, template, greedy decoding and budget).

Prompt text

Replace {} with the target sentence.

Classify whether the TARGET SENTENCE contains information about the target place that could help characterize its land use, land cover, or geographic environment from remote sensing, either directly or through observable proxies.

Return exactly one token: yes or no.

Answer yes for information about vegetation, agriculture, forests, water, soil or surface, terrain, buildings, settlements, infrastructure, transport networks, mining, managed land, or other human or natural features with a spatial or remotely detectable signature.

Answer no for information only about history, administration, people, events, demographics, economy, navigation, or activities with no meaningful land-use, land-cover, or remotely detectable implication.

Output only the lowercase token yes or no.

TARGET SENTENCE: {}

Aggregate scores

Per-model macro averages across languages. Per-language 95% intervals and paired tests: leaderboard.csv; full macro metrics: aggregates.csv. Bold = best; underline = second best in each metric column.

model_id	language_count	accuracy_macro	balanced_accuracy_macro	f1_macro	precision_macro	recall_macro	matthews_corrcoef_macro
Qwen/Qwen3.5-9B	85	0.7851	0.7775	0.8151	0.7529	0.8919	0.5786
LiquidAI/LFM2.5-2.6B@sglang-throughput-b16	85	0.7629	0.773	0.8106	0.7508	0.882	0.5636
LiquidAI/LFM2.5-2.6B+DSpark-throughput-b16	85	0.7623	0.7744	0.8104	0.7547	0.8763	0.5644
LiquidAI/LFM2.5-2.6B	85	0.7602	0.7536	0.8029	0.7253	0.9001	0.5352
Qwen/Qwen3-4B-Instruct-2507	85	0.7622	0.755	0.7905	0.7356	0.864	0.5344
Qwen/Qwen3-4B	85	0.7472	0.745	0.7634	0.7564	0.7784	0.4965
google/gemma-4-E4B-it	85	0.7728	0.78	0.7509	0.873	0.6719	0.5726
Qwen/Qwen3-8B	85	0.7553	0.7606	0.7442	0.831	0.6818	0.5283
Qwen/Qwen3.5-0.8B	85	0.6116	0.5882	0.7217	0.5883	0.939	0.2393
LiquidAI/LFM2.5-8B-A1B	85	0.7047	0.7056	0.7179	0.7318	0.7096	0.4128
LiquidAI/LFM2.5-8B-A1B+DSpark-throughput-b16	85	0.6995	0.7047	0.6992	0.7559	0.655	0.4118
LiquidAI/LFM2.5-350M	85	0.5333	0.5	0.6957	0.5333	1.0	0.0
Qwen/Qwen3-0.6B	85	0.5333	0.5	0.6957	0.5333	1.0	0.0
LiquidAI/LFM2.5-8B-A1B@sglang-throughput-b16	85	0.6936	0.6999	0.6915	0.753	0.644	0.4024
Qwen/Qwen3.5-4B	85	0.6833	0.6999	0.5936	0.9093	0.4515	0.4522
mistralai/Ministral-3-8B-Instruct-2512-BF16	85	0.6688	0.6869	0.5625	0.9164	0.4156	0.4344
google/gemma-4-E2B-it	85	0.6389	0.6577	0.5209	0.8746	0.3768	0.3732
LiquidAI/LFM2.5-1.2B-Instruct	85	0.6038	0.6158	0.516	0.7101	0.4351	0.2492
LiquidAI/LFM2.5-1.2B-Instruct+DSpark-throughput-b16	85	0.6024	0.6148	0.5095	0.7118	0.4274	0.2482
LiquidAI/LFM2.5-1.2B-Instruct@sglang-throughput-b16	85	0.6024	0.6148	0.5095	0.7118	0.4274	0.2482
LiquidAI/LFM2.5-VL-3B	85	0.6061	0.6239	0.4896	0.7951	0.3571	0.2912
LiquidAI/LFM2.5-VL-3B+DSpark-throughput-b16	85	0.6059	0.6237	0.4889	0.7951	0.3564	0.2909
LiquidAI/LFM2.5-VL-3B@sglang-throughput-b16	85	0.6059	0.6237	0.4889	0.7951	0.3564	0.2909
microsoft/Phi-4-mini-instruct	85	0.6026	0.6221	0.4598	0.8295	0.3297	0.3005
HuggingFaceTB/SmolLM3-3B	85	0.6071	0.6277	0.4551	0.8619	0.3187	0.3207
tiiuae/Falcon3-3B-Instruct	85	0.5909	0.6099	0.4432	0.7935	0.3261	0.2673
ibm-granite/granite-3.3-2b-instruct	85	0.5657	0.5846	0.4061	0.7213	0.3013	0.2016
unsloth/Qwen3.8-27B-GGUF@UD-IQ2_XXS	85	0.5863	0.6113	0.3664	0.9521	0.2363	0.321
allenai/Olmo-3-7B-Instruct	85	0.5688	0.5929	0.3469	0.8565	0.2315	0.2582
Qwen/Qwen3.5-2B	85	0.5613	0.5864	0.3306	0.8739	0.209	0.2569
mistralai/Ministral-3-3B-Instruct-2512-BF16	85	0.5561	0.5815	0.3173	0.8811	0.201	0.2507
utter-project/EuroLLM-9B-Instruct-2512	85	0.5276	0.5564	0.2125	0.9213	0.124	0.2129
tiiuae/Falcon3-7B-Instruct	85	0.5139	0.5438	0.1657	0.9132	0.095	0.1806
ibm-granite/granite-4.1-3b	85	0.5111	0.5415	0.1547	0.9729	0.0863	0.1864
allenai/OLMo-2-1124-7B-Instruct	85	0.5035	0.5341	0.1324	0.929	0.0754	0.1543
Qwen/Qwen3-1.7B	85	0.4893	0.5209	0.0843	0.7628	0.0468	0.1093
tiiuae/Falcon-H1-3B-Instruct	85	0.4878	0.5202	0.08	0.865	0.0428	0.1179
swiss-ai/Apertus-8B-Instruct-2509	85	0.4761	0.5089	0.0346	0.8809	0.0178	0.0813
tiiuae/Falcon3-1B-Instruct	85	0.4705	0.5035	0.0191	0.418	0.0099	0.0263

Scoring models

Scores are normalized to [0, 1]. Best thresholds are selected on this benchmark (an upper bound); ROC-AUC needs no threshold. Full sweep: threshold_sweep.csv.

Scoring setup

model	handling	relevance score / decision rule	sequence length (tokens)	dtype / batch / seed	revision
Alibaba-NLP/gte-multilingual-reranker-base	sequence classifier: prompt + sentence pair	sigmoid relevance logit; yes if score ≥ 0.5	8192	bfloat16; batch 16; seed 0	8215cf04918ba6f7b6a62bb44238ce2953d8831c
BalaRajesh1/mmbert-small-nli	NLI zero-shot pipeline: sentence premise + hypothesis	entailment probability; yes if score ≥ 0.5	512	bfloat16; batch 16; seed 0	2e7a7a1b86760ec00c5596f86e5b8bcb9ff8d9fb
LiquidAI/LFM2.5-2.6B@logprob	causal LM, no decoding: LLM prompt + chat turn, empty think block	first-token P(yes) vs P(no); argmax over the first-token yes/no log-probabilities	8192	bfloat16; batch 16; seed 0	654f9463ce32b05d0429d76fe1f580b27d4c1ac0
LiquidAI/LFM2.5-Encoder-350M	bidirectional masked-LM encoder: task prompt + one mask; 15 supported languages	yes/no masked-token logits; argmax over the masked-token yes/no logits	8192	bfloat16; batch 16; seed 0	b886781f7c6f10ca9b7096e21b83e30a073c2f39
MoritzLaurer/bge-m3-zeroshot-v2.0	NLI zero-shot pipeline: sentence premise + hypothesis	entailment probability; yes if score ≥ 0.5	512	bfloat16; batch 16; seed 0	9abf1c8aaeb82a2447809c20753ed0b106b76652
MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7	NLI zero-shot pipeline: sentence premise + hypothesis	entailment probability; yes if score ≥ 0.5	512	bfloat16; batch 16; seed 0	b5113eb38ab63efdd7f280f8c144ea8b13f978ce
Qwen/Qwen3-Reranker-0.6B	causal-LM reranker: manual yes/no reranker turn	yes/no next-token probability; argmax over the native yes/no scores	model-defined	bfloat16; batch 16; seed 0	e61197ed45024b0ed8a2d74b80b4d909f1255473
Qwen/Qwen3-Reranker-4B	causal-LM reranker: manual yes/no reranker turn	yes/no next-token probability; argmax over the native yes/no scores	model-defined	bfloat16; batch 16; seed 0	22e683669bc0f0bd69640a1354a6d0aebcfeede5
convaiinnovations/laya-multilingual	typed decision model: JSON state + 4 noul questions/call	Laya noul yes probability; yes if score ≥ 0.5	1024	varies across runs (dtype is recorded per run); batch 16; seed 0	b4a904d1a2a54c822b829e24291d4b8f280fe43e
fastino/gliner2.5-multi-v1	GLiNER2 classify_text: sentence + hypothesis label	label confidence; yes if score ≥ 0.5	model-defined	float32; batch 16; seed 0	a221b77a8baf4a613b8f8652661d41fa10a5641e
knowledgator/gliclass-multilang-mini	GLiClass zero-shot: sentence + hypothesis label	label probability; yes if score ≥ 0.5	512	bfloat16; batch 16; seed 0	0bd888b6c3ef9fca5f0a9d407bddfbbc7623486b
mixedbread-ai/mxbai-rerank-base-v2	causal-LM reranker: official query/document turn	sigmoid(1-logit - 0-logit - 4.5); yes if score ≥ 0.5	8192	bfloat16; batch 16; seed 0	3ea9d4dffa7d12a4f366be8e275c349de9fc9865

Scoring prompts

Alibaba-NLP/gte-multilingual-reranker-base, Qwen/Qwen3-Reranker-0.6B, Qwen/Qwen3-Reranker-4B, convaiinnovations/laya-multilingual, mixedbread-ai/mxbai-rerank-base-v2:

<Instruct>: Judge whether the Document carries information about its target place that could help characterize land use, land cover, or the geographic environment from remote sensing, either directly or through observable proxies. Answer yes for vegetation, agriculture, forests, water, soil or surface, terrain, buildings, settlements, infrastructure, transport networks, mining, managed land, or other human or natural features with a spatial or remotely detectable signature. Answer no for information only about history, administration, people, events, demographics, economy, navigation, or activities with no meaningful land-use, land-cover, or remotely detectable implication.
<Query>: Does this sentence carry land-use, land-cover, or geographic-environment signal observable from remote sensing?
<Document>: {}

BalaRajesh1/mmbert-small-nli, MoritzLaurer/bge-m3-zeroshot-v2.0, MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7, fastino/gliner2.5-multi-v1, knowledgator/gliclass-multilang-mini:

Zero-shot classification input for NLI and label-matching models. The premise is the TARGET SENTENCE; the hypothesis / positive label is the line after HYPOTHESIS, derived from the LLM prompt.

TARGET SENTENCE: {}

HYPOTHESIS: This sentence contains information useful for inferring the land use or land cover of the associated geographic place.

LiquidAI/LFM2.5-Encoder-350M:

Classify whether the target sentence contains information about its target place that
could help characterize land use, land cover, or geographic environment from features
observable by remote sensing.

Fill the blank with yes or no.

Answer yes for information about vegetation, agriculture, forests, water, soil or
surface, terrain, buildings, settlements, infrastructure, transport networks, mining,
managed land, or other human or natural features with a spatial or remotely detectable
signature.

Answer no for information only about history, administration, people, events,
demographics, economy, navigation, or activities with no meaningful land-use,
land-cover, or remotely detectable implication.

TARGET SENTENCE: {}
ANSWER: [MASK]

LiquidAI/LFM2.5-2.6B@logprob uses the task prompt above.

Best thresholded scoring metrics

model	languages	MCC @ threshold	F1 @ threshold	balanced accuracy @ threshold	precision @ threshold	recall @ threshold	ROC-AUC	items/s	peak VRAM (GiB)
LiquidAI/LFM2.5-2.6B@logprob	85	0.3942 @ 0.9	0.7307 @ 0.8	0.6884 @ 0.9	0.7109 @ 0.9	1 @ 0	0.7773	112.48	6.76
Qwen/Qwen3-Reranker-4B	85	0.2897 @ 0.03	0.723 @ 0.003	0.6435 @ 0.03	0.8288 @ 0.3	1 @ 0	0.7051	n/a	n/a
MoritzLaurer/bge-m3-zeroshot-v2.0	85	0.2573 @ 0.1	0.7041 @ 0.03	0.6226 @ 0.1	0.8344 @ 0.6	1 @ 0	0.6651	112.57	1.25
knowledgator/gliclass-multilang-mini	85	0.1987 @ 0.2	0.7037 @ 0.03	0.5976 @ 0.2	0.6383 @ 0.3	1 @ 0	0.6258	258.37	n/a
Qwen/Qwen3-Reranker-0.6B	85	0.187 @ 0.0003	0.7026 @ 1e-05	0.593 @ 0.0003	0.7863 @ 0.03	1 @ 0	0.6353	n/a	n/a
mixedbread-ai/mxbai-rerank-base-v2	85	0.1416 @ 0.003	0.7007 @ 0.001	0.5608 @ 0.01	0.6345 @ 0.01	1 @ 0	0.6153	14.34	7.14
LiquidAI/LFM2.5-Encoder-350M	15	0.0912 @ 0.9	0.6967 @ 0.8	0.5171 @ 0.9	0.5424 @ 0.9	1 @ 0	0.5462	194.86	1.75
convaiinnovations/laya-multilingual	85	0.0383 @ 0.4	0.6959 @ 0.01	0.5056 @ 0.4	0.5361 @ 0.4	1 @ 0	0.4871	156.37	1.55
Alibaba-NLP/gte-multilingual-reranker-base	85	0.0593 @ 0.5	0.6957 @ 0	0.5202 @ 0.6	0.5665 @ 0.6	1 @ 0	0.5409	78.64	1.29
BalaRajesh1/mmbert-small-nli	85	0 @ 0	0.6957 @ 0	0.5 @ 0	0.5333 @ 0	1 @ 0	0.4992	373.45	0.42
MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7	85	0.005 @ 0.9	0.6957 @ 0	0.5009 @ 0.9	0.5338 @ 0.9	1 @ 0	0.5284	260.91	0.74
fastino/gliner2.5-multi-v1	85	0.1305 @ 0.9	0.6957 @ 0	0.5607 @ 0.9	0.5833 @ 0.9	1 @ 0	0.6075	61.97	1.18

Observed output behavior: In these runs, LFM2.5-Encoder-350M predicted yes for 298 to 300 of 300 items per language across 15 languages.

LFM2.5-2.6B: log-probabilities vs generation

Same model, prompt and languages; log-probs call yes when P(yes) > P(no).

method	F1	MCC	unparsed	ROC-AUC	GPU hours	ms/item
generation + parsing	0.8029	0.5352	0.0042	n/a	40.11	5663.1
yes/no log-probs	0.6965	0.0191	0.0	0.7773	0.06	9.0
same GPU (NVIDIA A100-SXM4-40GB; en, fr, zh): generation vs log-probs					2462 s vs 7.1 s (349x)

Runtime performance

Timings are generation wall seconds summed across language runs; latency and throughput are recomputed from prediction telemetry. Different devices and runtimes are not directly comparable. Full per-language measurements are in leaderboard.csv.

Reproducibility probes show GPU sensitivity: on 15 overlapping LFM2.5-2.6B SGLang-throughput languages, A40 versus RTX 6000 Ada changed 474/4,500 verdicts; LFM2.5-8B-A1B changed 31/300 verdicts across GPU types. VL-3B SGLang throughput changed 158/25,500 verdicts between RTX A6000 and RTX 6000 Ada, while its same-GPU DSpark comparison changed 0/25,500. Throughput mode versus batch size 1 changed 2/300 VL-3B English verdicts. Transformers continuous batching fails on LFM2 with Invalid group type: conv. Compare runs only with the same runtime, mode and GPU model.

model_id	runtime	device	generation_mode	batch_size	language_count	cumulative_wall_seconds	sentences_per_second	latency_mean_seconds	latency_p50_seconds	latency_p95_seconds	generated_tokens	output_tokens_per_second	mean_accept_length	draft_accept_rate
Alibaba-NLP/gte-multilingual-reranker-base	transformers		static-batched	16	85	326.72	78.049
BalaRajesh1/mmbert-small-nli	transformers	Quadro RTX 6000	static-batched	16	85	71.8	355.143
HuggingFaceTB/SmolLM3-3B	transformers		static-batched	16	85	4506.36	5.659
LiquidAI/LFM2.5-1.2B-Instruct	transformers		static-batched	16	85	771.63	33.047
LiquidAI/LFM2.5-1.2B-Instruct+DSpark-throughput-b16	sglang	NVIDIA L40S	sglang-throughput	16	85	138.34	184.328	0.0783	0.0621	0.0997	51000	368.66	2.0	0.1111
LiquidAI/LFM2.5-1.2B-Instruct@sglang-throughput-b16	sglang	varies	sglang-throughput	16	85	116.89	218.156	0.0651	0.0539	0.0884	51000	436.31
LiquidAI/LFM2.5-2.6B	transformers		static-batched	16	85	144409.53	0.177
LiquidAI/LFM2.5-2.6B+DSpark-throughput-b16	sglang	varies	sglang-throughput	16	85	17660.83	1.444	4.3867	3.3135	11.0612	28688903	1624.44	3.965	0.3291
LiquidAI/LFM2.5-2.6B@logprob	transformers	NVIDIA A100-SXM4-40GB	static-batched	16	85	230.01	110.864
LiquidAI/LFM2.5-2.6B@sglang-throughput-b16	sglang	varies	sglang-throughput	16	85	38332.09	0.665	9.1619	7.0569	24.1868	28575570	745.47
LiquidAI/LFM2.5-350M	transformers		static-batched	16	85	353.71	72.092
LiquidAI/LFM2.5-8B-A1B	transformers		static-batched	16	85	118305.81	0.216
LiquidAI/LFM2.5-8B-A1B+DSpark-throughput-b16	sglang	varies	sglang-throughput	16	85	11027.38	2.312	3.7921	3.365	7.2036	10491590	951.41	4.093	0.3425
LiquidAI/LFM2.5-8B-A1B@sglang-throughput-b16	sglang	varies	sglang-throughput	16	85	16132.6	1.581	5.4952	4.8097	10.2668	10543669	653.56
LiquidAI/LFM2.5-Encoder-350M	transformers	NVIDIA L4	static-batched	16	15	24.88	180.883
LiquidAI/LFM2.5-VL-3B	transformers	NVIDIA L40S	static-batched	16	85	248.13	102.769	0.1539	0.1479	0.1879	51000	205.54
LiquidAI/LFM2.5-VL-3B+DSpark-throughput-b16	sglang	NVIDIA RTX 6000 Ada Generation	sglang-throughput	16	85	335.76	75.948	0.2011	0.1995	0.239	51000	151.9	2.0	0.3362
LiquidAI/LFM2.5-VL-3B@sglang-throughput-b16	sglang	NVIDIA RTX 6000 Ada Generation	sglang-throughput	16	85	291.37	87.516	0.1737	0.1731	0.2091	51000	175.03
MoritzLaurer/bge-m3-zeroshot-v2.0	transformers	Quadro RTX 6000	static-batched	16	85	228.53	111.583
MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7	transformers	Quadro RTX 6000	static-batched	16	85	98.72	258.312
Qwen/Qwen3-0.6B	transformers		static-batched	16	85	280.58	90.883
Qwen/Qwen3-1.7B	transformers		static-batched	16	85	537.17	47.471
Qwen/Qwen3-4B	transformers		static-batched	16	85	908.71	28.062
Qwen/Qwen3-4B-Instruct-2507	transformers		static-batched	16	85	829.5	30.741
Qwen/Qwen3-8B	transformers		static-batched	16	85	890.93	28.622
Qwen/Qwen3-Reranker-0.6B	transformers		static-batched	16	85	177.81	143.41
Qwen/Qwen3-Reranker-4B	transformers		static-batched	16	85	662.51	38.49
Qwen/Qwen3.5-0.8B	transformers		static-batched	16	85	321.1	79.414
Qwen/Qwen3.5-2B	transformers		static-batched	16	85	377.11	67.62
Qwen/Qwen3.5-4B	transformers		static-batched	16	85	1096.92	23.247
Qwen/Qwen3.5-9B	transformers		static-batched	16	85	1614.21	15.797
allenai/OLMo-2-1124-7B-Instruct	transformers		static-batched	16	85	915.11	27.866
allenai/Olmo-3-7B-Instruct	transformers		static-batched	16	85	1067.17	23.895
convaiinnovations/laya-multilingual	transformers		static-batched	16	85	180.29	141.44
fastino/gliner2.5-multi-v1	transformers	NVIDIA A100-SXM4-40GB	static-batched	16	85	411.7	61.939
google/gemma-4-E2B-it	transformers		static-batched	16	85	430.29	59.262
google/gemma-4-E4B-it	transformers		static-batched	16	85	817.29	31.201
ibm-granite/granite-3.3-2b-instruct	transformers		static-batched	16	85	547.74	46.555
ibm-granite/granite-4.1-3b	transformers		static-batched	16	85	437.63	58.269
knowledgator/gliclass-multilang-mini	transformers	Quadro RTX 6000	static-batched	16	85	100.12	254.699
microsoft/Phi-4-mini-instruct	transformers		static-batched	16	85	563.83	45.226
mistralai/Ministral-3-3B-Instruct-2512-BF16	transformers		static-batched	16	85	1707.26	14.936
mistralai/Ministral-3-8B-Instruct-2512-BF16	transformers		static-batched	16	85	1455.05	17.525
mixedbread-ai/mxbai-rerank-base-v2	transformers		static-batched	16	85	1880.78	13.558
swiss-ai/Apertus-8B-Instruct-2509	transformers		static-batched	16	85	1300.16	19.613
tiiuae/Falcon-H1-3B-Instruct	transformers		static-batched	1	85	7697.51	3.313
tiiuae/Falcon3-1B-Instruct	transformers		static-batched	16	85	278.09	91.696
tiiuae/Falcon3-3B-Instruct	transformers		static-batched	16	85	427.21	59.69
tiiuae/Falcon3-7B-Instruct	transformers		static-batched	16	85	910.7	28.0
unsloth/Qwen3.8-27B-GGUF@UD-IQ2_XXS	transformers	NVIDIA A100-SXM4-40GB	static-batched	16	85	6530.99	3.904
utter-project/EuroLLM-9B-Instruct-2512	transformers		static-batched	16	85	673.97	37.835

DSpark speculative decoding

Greedy DSpark runs should match the same-target SGLang baseline across every language; the check requires complete item coverage and identical generated text.

target_model	sglang_baseline	dspark_run	baseline_output_tokens_per_second	dspark_output_tokens_per_second	speedup	identical_predictions	languages_compared	items_compared	verdict_differences	text_differences
LiquidAI/LFM2.5-1.2B-Instruct	LiquidAI/LFM2.5-1.2B-Instruct@sglang-throughput-b16	LiquidAI/LFM2.5-1.2B-Instruct+DSpark-throughput-b16	436.31	368.66	0.84x	yes	85	25500	0	0
LiquidAI/LFM2.5-2.6B	LiquidAI/LFM2.5-2.6B@sglang-throughput-b16	LiquidAI/LFM2.5-2.6B+DSpark-throughput-b16	745.47	1624.44	2.18x	no	85	25500	3004	24962
LiquidAI/LFM2.5-8B-A1B	LiquidAI/LFM2.5-8B-A1B@sglang-throughput-b16	LiquidAI/LFM2.5-8B-A1B+DSpark-throughput-b16	653.56	951.41	1.46x	no	85	25500	4641	25029
LiquidAI/LFM2.5-VL-3B	LiquidAI/LFM2.5-VL-3B@sglang-throughput-b16	LiquidAI/LFM2.5-VL-3B+DSpark-throughput-b16	175.03	151.9	0.87x	yes	85	25500	0	0