Pith. sign in

REVIEW 3 major objections 4 minor 44 references

Cost-Optimal Foundation Model Deployment Portfolio for Transportation Management

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Foundation-model deployment for traffic centers is a portfolio problem, not a set of independent choices, and the optimal mix can cut monthly costs by 97 percent versus the cheapest feasible all-closed-API baseline.

desk verdict A clean MIP formulation for multi-function FM deployment with a valid NP-hardness proof, but the headline 97% saving is an illustrative artifact of unmeasured FNR inputs, not an empirical result. read the letter →

arxiv 2607.13239 v1 pith:DOMI4OWQ submitted 2026-07-14 cs.AI

Xi Cheng , Ke Liu , Siyuan Feng , Jane Lin , H. Oliver Gao This is my paper
classification cs.AI MSC 90C1090C27
keywords foundationmodeldeploymentmixed-integerprogrammingtotalcostofownershiptransportationmanagementcenterGPUcapacityNP-hardnessportfoliooptimizationbreak-evenanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that choosing which foundation model serves each transportation-management function should be a portfolio decision, not a set of independent choices. It formulates FMDP, a mixed-integer program that minimizes monthly total cost of ownership subject to per-function quality, latency, safety, and shared GPU-memory constraints. It proves FMDP NP-hard by reduction from 0-1 knapsack and gives a polynomial-time greedy heuristic. In an illustrative five-function case study, the optimal portfolio routes four functions to hosted open-source APIs and one function to a closed API because no open-source model meets that function's quality floor, yielding $34/month versus $1,110/month for the cheapest feasible all-closed baseline. The paper also derives query-rate break-even thresholds that tell agencies when owning GPUs beats paying for APIs.

What carries the argument

The central object is the FMDP mixed-integer program: binary assignment variables x_fmd choose a (model, mode) pair per function, binary loading variables y_md capture fixed-charge GPU costs, and the objective sums API token costs, fixed hardware costs, and energy costs. Constraints enforce quality floors, latency ceilings, safety bounds for critical functions, and GPU capacity per mode. NP-hardness comes from a reduction of 0-1 knapsack to FMDP via per-function local models with memory weights and API fallbacks with engineered costs. The polynomial greedy heuristic does a cheap initial assignment, then repairs GPU-capacity violations by switching the function with the smallest repair ratio

What would settle it

Run the five TMC functions on a benchmark with the actual candidate models and measure quality, false-negative rate, and latency on TMC data. If any open-source model achieves at least 0.80 on incident report generation while meeting latency and safety constraints, the optimal FMDP portfolio would shift and the 97 percent gap versus all-closed baselines would shrink or disappear.

Watch

Extended reading notes

Core claim

The paper's central claim is that a mixed deployment is both feasible and dramatically cheaper than any single-tier strategy. FMDP assigns each TMC function exactly one (model, mode) pair while satisfying per-function quality, latency, and safety constraints and a shared GPU capacity limit. The cost model separates per-query API token costs from fixed hardware charges and variable energy costs, so loading a model on a GPU is a fixed charge shared across functions. The case study finds a $34/month portfolio that is feasible because it reserves the only closed-API slot for incident report generation, where the quality floor is 0.80 and the best open-source candidate scores 0.78. The break-even

Load-bearing premise

The 97 percent cost gap rests on the assumption, stated as illustrative, that no open-source model reaches the quality floor q_min=0.80 for incident report generation while GPT-4o-mini (q=0.82) does; if actual TMC measurements differ, the optimal portfolio and the claimed gap change.

Editorial extensions

If this is right

  • Agencies should evaluate foundation-model deployment as a portfolio with a shared GPU budget rather than per-function independent choices.
  • An all-open-source API strategy can be nearly as cheap but infeasible when at least one function has a quality floor no open-source model meets; one targeted closed-API slot fixes feasibility at near-zero marginal cost.
  • On-premise GPUs are not cost-effective for low-query text workloads in this scenario; the break-even rate for a vision API is about 309 queries/hour on a single RTX 4090 at current prices.
  • If vision API prices double, the break-even rate falls to about 154 queries/hour, flipping f1's optimal deployment from API to on-premise.
  • The greedy heuristic runs in polynomial time, so agencies can re-run the portfolio whenever price lists or model catalogs change.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The optimization structure is not specific to traffic centers: any organization choosing among cloud APIs, hosted open-source models, and self-hosted GPUs for multiple AI functions with a shared hardware budget faces the same knapsack-with-fixed-charge structure, so the NP-hardness result and greedy repair logic carry over.
  • The break-even formula lambda* = C_hw/(H*c_q) gives a simple pre-purchase rule the paper does not emphasize: compute lambda* per function, and only functions whose query rate exceeds their threshold justify local hardware.
  • Because FMDP assumes exactly one model per function, the $34/month figure is an upper bound on the minimal achievable cost under the stated catalog; allowing per-query cascading or fine-tuned open-source models could only lower or keep the cost equal, subject to the same constraints.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper formulates the Foundation Model Deployment Portfolio (FMDP) problem: given a set of TMC functions with quality, latency, and safety requirements, choose one feasible (model, deployment mode) pair per function to minimize monthly TCO under shared GPU-memory constraints. The authors present a mixed-integer program (Section III.C), prove NP-hardness by reduction from 0-1 knapsack (Section IV.A), propose a two-phase greedy heuristic (Section IV.C), and give an illustrative five-function, 19-candidate case study (Section V). The headline result is a $34/mo mixed portfolio, 97% below the cheapest feasible all-closed-API baseline, with break-even analysis for on-premise vs. API deployment.

Significance. The contribution is a useful formalization of an emerging practical problem. The NP-hardness proof is a genuine reduction from 0-1 knapsack and is internally sound; the MIP is clearly specified; the greedy heuristic is plausible and polynomial-time; and Eq. (12) gives a simple, actionable break-even rule. If the framework is the main contribution, it is a reasonable one. However, the quantitative 97% claim is entirely dependent on hand-assigned, unmeasured quality and FNR scores, and the case study does not report the latency values that Eq. (8) requires. The paper is more convincing as a modeling and complexity contribution than as an empirical demonstration of cost savings.

major comments (3)
  1. [Section V.A / Table IV] The central quantitative claim—a $34/mo portfolio, 97% below the all-closed baseline—rests on the unmeasured FNR and quality scores assigned to f1. Table IV gives Qwen2.5-VL-7B FNR=0.08 and q=0.78, allowing f1 (144,000 queries/month, $1,102 via GPT-4o) to be served for $29/mo. The text states these values are 'illustrative' and that the authors 'assume that the relative performance of these models on TMC tasks mirrors their ranking on public benchmarks' (Section V.A), but no measurement or citation substantiates FNR=0.08 for wrong-way/pedestrian detection. If the true FNR exceeds 0.10, no open-source candidate in Table IV satisfies f1's safety constraint, and the optimal cost jumps to roughly $1,107/mo, eliminating the 97% gap. Because the abstract presents the 97% figure as a key result, the case study must either supply measured FNR/quality values or be reframed explicitly as a hypothe
  2. [Section V / Table IV / Eq. (8)] Table IV lists no latency values ℓ_fmd for any candidate, even though Eq. (8) imposes a per-function latency ceiling L_f (Table III). The abstract claims the FMDP portfolio satisfies latency constraints, but the case study never instantiates this constraint. For example, f1 has L=5 s and a 200 q/h rate, yet Table IV gives no indication of whether GPT-4o or Qwen2.5-VL API calls, or on-premise inference, actually meet 5 s. Without these values, the feasibility of FMDP and of every baseline cannot be verified. The authors should add a latency column (or at minimum explicitly state an assumption, with evidence, that all listed candidates satisfy the ceilings). This is a required fix, not a cosmetic one.
  3. [Table IV / Table V] The baselines in Table V are not reproducible from the 'full catalog' in Table IV. All-On-Premise is reported as requiring 63 GB, but Table IV lists on-premise options only for f1 (InternVL2-8B, 16 GB) and f2 (Qwen2.5-7B, 15 GB); f3, f4, and f5 have no on-premise rows. Similarly, Classical-Only lacks specified candidates for f2, f4, and f5. The statement 'Table IV lists the full catalog' is therefore inconsistent with the baseline construction. The authors should either extend Table IV to include all candidates used by the baselines or restrict the baselines to the catalog actually shown.
minor comments (4)
  1. [Section V.B / Table VI] Eq. (12) with C_hw=$44/mo, H=720, c_q=$0.0002 (Qwen2.5-VL) gives λ* = 44/(720×0.0002) ≈ 305.6 q/h, not the 308.6 shown in Table VI and 309 in the text. Please reconcile the arithmetic, e.g., by specifying the exact c_q or hardware cost used.
  2. [Section IV.B] The claim that the per-function LP relaxation is integral is correct only after infeasible (m,d) pairs are pre-filtered so that constraints (6)–(8) are redundant. As written, the sentence could be misread as applying to the original constraints; clarify that the integrality statement relies on the pre-filtering described in Section III.C.
  3. [Section V.A] The sentence 'We assume that the relative performance of these models on TMC tasks mirrors their ranking on public benchmarks' is a strong assumption but is not accompanied by any benchmark citation for the specific FNR or quality numbers. At minimum, point to the exact source for each score or label the scores as expert guesses.
  4. [Section IV.C] In Phase 2 of the greedy, when a function shares its GPU model with others, the text sets g_f=0 and r_f=∞; if all r_f are ∞, the heuristic switches the function with the smallest Δ_f. This behavior is described correctly, but a short example or proof of termination would help the reader see that the loop always reduces G_used eventually.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: cost model, MIP, NP-hardness proof, and break-even formulas are self-contained; the case study is explicitly illustrative rather than a prediction.

full rationale

The paper's derivation chain does not reduce to its own inputs. The FMDP formulation (Eqs. 1-11) is a standard mixed-integer program built from public API prices and an explicitly illustrative catalog; the 'optimal portfolio' is the optimizer's output for those given parameters, not an independent empirical prediction. The NP-hardness proof reduces from 0-1 knapsack by constructing an FMDP instance with two candidates per function and non-binding quality/latency/safety constraints, which genuinely shows hardness rather than assuming it. The break-even formula (Eq. 12) is algebraically derived from the cost model: setting hardware cost equal to per-query API cost gives λ* = C_hw/(H·c_q), with no fitted constants. The hand-assigned quality and FNR values in Table IV are acknowledged as illustrative in both Section V.A ('illustrative but align with published performance tiers') and the Limitations section ('quality, latency, and safety parameters are illustrative and should be replaced with agency-specific measurements'), so the 97% case-study figure is honestly presented as a scenario calculation, not as a measured prediction. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The only weakness is external validity of the illustrative parameters, which is a correctness/evidence concern, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The MIP core and hardness proof are self-contained; the case study pulls its empirical content from hand-assigned quality/FNR values and an untested benchmark-to-TMC transfer assumption. The $34/mo result is therefore a demonstration, not a measurement.

free parameters (4)
  • Quality scores q_fm = e.g., f3: Llama-3.1-8B q=0.78 vs q_min=0.80; GPT-4o-mini q=0.82
    Hand-assigned 'illustrative' values in Table IV, asserted to mirror public benchmarks; they determine which portfolios are feasible and hence the 97% figure.
  • FNR values s_f1m = GPT-4o .04, Qwen2.5-VL .08, InternVL .08, YOLOv8 .11
    Illustrative safety metrics for f1; YOLO is excluded by the FNR<=0.10 bound. No measurements on TMC CCTV data.
  • Candidate catalog composition A_f = 19 (model, mode) pairs in Table IV
    Which models/modes are offered per function is a modeling choice; no open-source candidate for f3 meets q_min=0.80 by construction, and no larger open-source LLM is included.
  • Latency values ℓ_fmd = not reported
    Latency constraints are in the formulation but no per-pair latencies appear in Table IV; feasibility w.r.t. latency is assumed by omission.
assumptions (4)
  • domain assumption Model performance on TMC tasks mirrors public benchmark rankings (quality and FNR)
    Section V.A: 'We assume that the relative performance of these models on TMC tasks mirrors their ranking on public benchmarks.' Without this, the quality/FNR inputs are unsupported.
  • domain assumption Safety for video anomaly detection can be captured by a single model-level FNR bound
    Section III.A introduces s_fm <= S_max with FNR as a natural choice; no operational validation is provided.
  • domain assumption GPU utilization for energy cost is sequential: μ_fmd = min(λ_f ℓ_fmd/3600, 1)
    Section III.B; this ignores batching and concurrency, affecting on-premise energy cost and break-even thresholds.
  • standard math 0-1 knapsack is NP-hard (Karp)
    Used as the basis of the reduction in Proposition 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cost-Optimal Foundation Model Deployment Portfolio for Transportation Management." pith.science (2026). https://pith.science/paper/DOMI4OWQ

@misc{pith2026260713239,
  author       = {Pith},
  title        = {Pith review of: Cost-Optimal Foundation Model Deployment Portfolio for Transportation Management},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DOMI4OWQ}},
  note         = {Machine review of arXiv:2607.13239}
}
abstract

Foundation models, including large language models (LLMs) and vision-language models (VLMs), are increasingly used for transportation management center (TMC) tasks such as anomaly detection, incident reporting, and traveler information. Deploying multiple such models across TMC functions raises a portfolio question: which model should serve each function, in which deployment mode, and under what shared hardware budget? We formulate this as the Foundation Model Deployment Portfolio (FMDP) problem, a mixed-integer program minimizing total cost of ownership (TCO) subject to per-function quality, latency, and safety constraints over shared GPU capacity. We prove the problem NP-hard by reduction from the 0-1 knapsack problem and propose a polynomial-time greedy heuristic. In an illustrative case study with five TMC functions and 19 candidate (model, mode) pairs, FMDP identifies a mixed portfolio costing $34/mo (97% below the cheapest feasible all-closed-API baseline) by routing four functions to open-source APIs and the one function whose quality floor no open-source model meets to a closed API. Break-even analysis shows that on-premise GPU investment becomes reasonable only above approximately 309 vision queries/hour or if API prices double.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 3 linked inside Pith

  1. [1]

    Next generation of traffic manage- ment systems and centers: A primer,

    Federal Highway Administration, “Next generation of traffic manage- ment systems and centers: A primer,” tech. rep., U.S. Department of Transportation, 2024

    2024

  2. [2]

    Next generation transportation management centers and transportation management systems: Executive briefing,

    ITS Joint Program Office, “Next generation transportation management centers and transportation management systems: Executive briefing,” tech. rep., U.S. Department of Transportation, 2025. Updated January 2026

    2025

  3. [3]

    Ultralytics YOLOv8

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics YOLOv8.” https: //github.com/ultralytics/ultralytics, 2023

    2023

  4. [4]

    Harnessing large language models for training-free video anomaly detection,

    L. Zanella, W. Menapace, M. Mancini, Y . Wang, and E. Ricci, “Harnessing large language models for training-free video anomaly detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 18527–18536, 2024

    2024

  5. [5]

    To- wards explainable traffic flow prediction with large language models,

    X. Guo, Q. Zhang, J. Jiang, M. Peng, M. Zhu, and H. F. Yang, “To- wards explainable traffic flow prediction with large language models,” Communications in Transportation Research, vol. 4, p. 100150, 2024

    2024

  6. [6]

    Open-ti: Open traffic intelligence with augmented language model,

    L. Da, K. Liou, T. Chen, X. Zhou, X. Luo, Y . Yang, and H. Wei, “Open-ti: Open traffic intelligence with augmented language model,” International Journal of Machine Learning and Cybernetics, vol. 15, no. 10, pp. 4761–4786, 2024

    2024

  7. [7]

    Llmlight: Large language models as traffic signal control agents,

    S. Lai, Z. Xu, W. Zhang, H. Liu, and H. Xiong, “Llmlight: Large language models as traffic signal control agents,” inProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 1, pp. 2335–2346, 2025

    2025

  8. [8]

    Large language models for intelligent transportation: A review of the state of the art and challenges,

    S. Wandelt, C. Zheng, S. Wang, Y . Liu, and X. Sun, “Large language models for intelligent transportation: A review of the state of the art and challenges,”Applied Sciences, vol. 14, no. 17, p. 7455, 2024

    2024

Show all 44 references
  1. [9]

    Large language models for mobility analysis in transportation systems: A survey on forecasting tasks,

    Z. Zhang, Y . Sun, Z. Wang, Y . Nie, X. Ma, R. Li, P. Sun, and X. Ban, “Large language models for mobility analysis in transportation systems: A survey on forecasting tasks,”Transportation Research Record, vol. 2680, no. 2, pp. 756–774, 2026

    2026

  2. [10]

    Multimodal perception and decision-making systems for complex roads based on foundation models,

    L. Fan, Y . Wang, H. Zhang, C. Zeng, Y . Li, C. Gou, and H. Yu, “Multimodal perception and decision-making systems for complex roads based on foundation models,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 54, no. 11, pp. 6561–6569, 2024

    2024

  3. [11]

    API pricing

    OpenAI, “API pricing.” https://openai.com/api/pricing, 2025. Ac- cessed: Jan. 2026

    2025

  4. [12]

    Gemini API pricing

    Google, “Gemini API pricing.” https://ai.google.dev/pricing, 2025. Accessed: Jan. 2026

    2025

  5. [13]

    The llama 3 herd of models,

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al- Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan,et al., “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  6. [14]

    Inference pricing

    Together AI, “Inference pricing.” https://www.together.ai/pricing,

  7. [15]

    Frugalgpt: How to use large language models while reducing cost and improving performance,

    L. Chen, M. Zaharia, and J. Zou, “Frugalgpt: How to use large language models while reducing cost and improving performance,” arXiv preprint arXiv:2305.05176, 2023

    2023 arXiv

  8. [16]

    Routellm: Learning to route llms from preference data,

    I. Ong, A. Almahairi, V . Wu, W.-L. Chiang, T. Wu, J. E. Gonzalez, M. W. Kadous, and I. Stoica, “Routellm: Learning to route llms from preference data,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  9. [17]

    Thriftllm: On cost-effective selection of large language models for classification queries,

    K. Huang, Y . Shi, D. Ding, Y . Li, Y . Fei, L. Lakshmanan, and X. Xiao, “Thriftllm: On cost-effective selection of large language models for classification queries,”Proceedings of the VLDB Endowment, vol. 18, no. 11, pp. 4410–4423, 2025

    2025

  10. [18]

    {INFaaS}: Automated model-less inference serving,

    F. Romero, Q. Li, N. J. Yadwadkar, and C. Kozyrakis, “{INFaaS}: Automated model-less inference serving,” in2021 USENIX Annual Technical Conference (USENIX ATC 21), pp. 397–411, 2021

    2021

  11. [19]

    Dota: Unsupervised detection of traffic anomaly in driving videos,

    Y . Yao, X. Wang, M. Xu, Z. Pu, Y . Wang, E. Atkins, and D. J. Crandall, “Dota: Unsupervised detection of traffic anomaly in driving videos,”IEEE transactions on pattern analysis and machine intelli- gence, vol. 45, no. 1, pp. 444–459, 2022

    2022

  12. [20]

    Hybrid llm: Cost-efficient and quality-aware query routing,

    D. Ding, A. Mallick, C. Wang, R. Sim, S. Mukherjee, V . R ¨uhle, L. V . Lakshmanan, and A. H. Awadallah, “Hybrid llm: Cost-efficient and quality-aware query routing,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  13. [21]

    A unified approach to routing and cascading for llms,

    J. Dekoninck, M. Baader, and M. Vechev, “A unified approach to routing and cascading for llms,” inICML 2025, 2025

    2025

  14. [22]

    Fly-swat or cannon? cost- effective language model choice via meta-modeling,

    M. ˇSakota, M. Peyrard, and R. West, “Fly-swat or cannon? cost- effective language model choice via meta-modeling,” inProceedings of the 17th ACM International Conference on Web Search and Data Mining, pp. 606–615, 2024

    2024

  15. [23]

    Cloud or on-premise? a strategic view of large language model deployment,

    Z. J. Zhang, J. Shi, and S. Tang, “Cloud or on-premise? a strategic view of large language model deployment,”A Strategic View of Large Language Model Deployment (June 16, 2025), 2025

    2025

  16. [24]

    A middle path for on-premises llm deployment: Preserving privacy without sacrificing model confidentiality,

    H. Huang, Y . Li, B. Jiang, B. Jiang, L. Liu, Z. Liu, R. Sun, and S. Liang, “A middle path for on-premises llm deployment: Preserving privacy without sacrificing model confidentiality,” inProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, p...

    2025

  17. [25]

    A multiplier adjustment method for the generalized assignment problem,

    M. L. Fisher, R. Jaikumar, and L. N. Van Wassenhove, “A multiplier adjustment method for the generalized assignment problem,”Manage- ment science, vol. 32, no. 9, pp. 1095–1103, 1986

    1986

  18. [26]

    Assignment problems: A golden anniversary survey,

    D. W. Pentico, “Assignment problems: A golden anniversary survey,” European Journal of Operational Research, vol. 176, no. 2, pp. 774– 793, 2007

    2007

  19. [27]

    M\’elange: Cost efficient large language model serving by exploiting gpu heterogeneity,

    T. Griggs, X. Liu, J. Yu, D. Kim, W.-L. Chiang, A. Cheung, and I. Stoica, “M\’elange: Cost efficient large language model serving by exploiting gpu heterogeneity,”arXiv preprint arXiv:2404.14527, 2024

    2024 arXiv

  20. [28]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,

    L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qin,et al., “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,”ACM Transactions on Information Systems, vol. 43, no. 2, pp. 1–55, 2025

    2025

  21. [29]

    A cost-benefit analysis of on-premise large language model deployment: Breaking even with commercial llm services,

    G. Pan, V . Chodnekar, A. Roy, and H. Wang, “A cost-benefit analysis of on-premise large language model deployment: Breaking even with commercial llm services,”arXiv preprint arXiv:2509.18101, 2025

    2025

  22. [30]

    LLM total cost of ownership 2025: Build vs. buy math

    O. Gubanova, “LLM total cost of ownership 2025: Build vs. buy math.” Ptolemay Blog, 2025. Accessed: Jan. 2026

    2025

  23. [31]

    Cost of serving LLMs

    S. Ivanovici, “Cost of serving LLMs.” Fin.ai Research Blog, 2025. Accessed: Jan. 2026

    2025

  24. [32]

    Cost comparison: API vs self-hosting for open-weight LLMs

    DetectX, “Cost comparison: API vs self-hosting for open-weight LLMs.” DetectX Blog, 2025. Accessed: Jan. 2026

    2025

  25. [33]

    The complete AI strategy guide: Cloud APIs vs self-hosted models

    T. Sharma, “The complete AI strategy guide: Cloud APIs vs self-hosted models.” Medium, 2025. Accessed: Jan. 2026

    2025

  26. [34]

    Efficient memory management for large language model serving with pagedattention,

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Yu, J. Gonzalez, H. Zhang, and I. Stoica, “Efficient memory management for large language model serving with pagedattention,” inProceedings of the 29th symposium on operating systems principles, pp. 611–626, 2023

    2023

  27. [35]

    Sglang: Efficient execution of structured language model programs,

    L. Zheng, L. Yin, Z. Xie, C. L. Sun, J. Huang, C. H. Yu, S. Cao, C. Kozyrakis, I. Stoica, J. E. Gonzalez,et al., “Sglang: Efficient execution of structured language model programs,”Advances in neural information processing systems, vol. 37, pp. 62557–62583, 2024

    2024

  28. [36]

    API pricing

    Anthropic, “API pricing.” https://claude.com/pricing#api, 2025. Ac- cessed: Jan. 2026

    2025

  29. [37]

    Qwen2.5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y . Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y . Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin, “Qwen2.5-vl technical report,” 2025

    2025

  30. [38]

    GPU cloud pricing

    RunPod, “GPU cloud pricing.” https://www.runpod.io/pricing, 2025. Accessed: Jan. 2026

    2025

  31. [39]

    GeForce RTX 4090

    NVIDIA, “GeForce RTX 4090.” https://www.nvidia.com/en-us/ geforce/graphics-cards/40-series/rtx-4090/, 2022. MSRP $1,599 at launch (Oct. 2022)

    2022

  32. [40]

    NVIDIA A100 tensor core GPU

    NVIDIA, “NVIDIA A100 tensor core GPU.” https://www.nvidia.com/ en-us/data-center/a100/, 2020. 80 GB PCIe; market price approxi- mately $15,000 (2025)

    2020

  33. [41]

    Reducibility among combinatorial problems,

    R. M. Karp, “Reducibility among combinatorial problems,” in50 Years of Integer Programming 1958-2008: from the Early Years to the State- of-the-Art, pp. 219–241, Springer, 2009

    1958

  34. [42]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks,

    Z. Chen, J. Wu, W. Wang, W. Su, G. Chen, S. Xing, M. Zhong, Q. Zhang, X. Zhu, L. Lu,et al., “Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, ...

    2024

  35. [43]

    Judging LLM-as-a-judge with MT-Bench and Chatbot Arena,

    L. Zheng, W.-L. Chiang, Y . Sheng,et al., “Judging LLM-as-a-judge with MT-Bench and Chatbot Arena,” inNeurIPS, 2023

    2023

  36. [44]

    Llm leaderboard

    Artificial Analysis, “Llm leaderboard.” https://artificialanalysis.ai/ leaderboards/models, 2025. Accessed: Jan. 2026

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.