Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Medicine on the Edge: Comparative Performance Analysis of On-Device LLMs for Clinical Reasoning

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Compact quantized language models can already produce clinically meaningful answers to open-ended medical cases on iPhones and iPads, and the main limit to wider deployment is device memory, not processing power.

desk verdict A genuinely useful first map of what runs on Apple devices for clinical QA, but the accuracy rankings rest on unvalidated protocol modifications and should be read as suggestive, not load-bearing. read the letter →

arxiv 2502.08954 v1 pith:XVOKXQKO submitted 2025-02-13 cs.CL

Leon Nissen , Philipp Zagar , Vishnu Ravi , Aydin Zahedivash , Lara Marie Reimer , Stephan Jonas , Oliver Aalami , Paul Schmiedmayer This is my paper · ORCID
classification cs.CL
keywords on-deviceLLMclinicalreasoningAMEGAbenchmarkquantizationmobilehealthMLXAppleSiliconedgeinference
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

This paper tries to establish that modern small language models—including some fine-tuned for medicine—can run entirely on a smartphone or tablet and still produce useful answers to open-ended clinical cases. The authors benchmark 13 open-weight models, quantized to 4-bit and run through the MLX framework, on six iOS devices using the AMEGA dataset of 20 patient vignettes. They report that the medically fine-tuned Med42 and Aloe 8B models reach the highest scores, about 490 of 1,000 points, while the much smaller general-purpose Phi-3 Mini scores 464.6 and gives the best speed-accuracy trade-off. They also find that even a 2020 iPhone can run smaller models and that the binding constraint is the operating system's memory limit, not raw processing power. If these results hold, privacy-preserving on-device clinical decision support is feasible on hardware people already own.

What carries the argument

The machinery is a mobile benchmarking harness rather than a single mathematical object: a custom iOS app built around the MLX framework, an array-based inference stack for Apple silicon, runs 4-bit-quantized models entirely on-device, clears the GPU cache after every question, and logs tokens per second, time to first token, memory use, thermal state, and battery level. The scored output comes from the AMEGA benchmark, whose 20 open-ended clinical cases are rated by a GPT-4o judge under a modified protocol: temperature 0.1 instead of 0, the follow-up reask step removed, and reasoning traces stripped before evaluation. The quantized MLX conversion is what lets 8-billion-parameter models fit into the 6 to 8 GB of memory on recent iPhones, while each device's iOS memory budget determines which models can run at all.

What would settle it

Run the same 13 quantized models through the original, unmodified benchmark procedure—temperature 0.0, follow-up reask enabled, reasoning text retained—and check whether the accuracy ranking changes materially; if Med42 or Aloe no longer lead, or Phi-3 Mini drops well below Llama 3.1 8B, the headline accuracy results are an artifact of the modified judging setup. A blind human clinician rating of the same responses would independently settle whether the GPT-4o scores reflect clinical quality.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a measured feasibility result: after 4-bit quantization and conversion to the MLX format, open-weight language models from 1 to 8 billion parameters can run on unmodified iPhones and iPads released between 2020 and 2024 and can answer AMEGA clinical cases at 8 to 120 tokens per second. Accuracy is led by the medical fine-tunes Aloe 8B (490.9) and Med42 8B (490.0), with the compact general-purpose Phi-3 Mini 3.8B close behind at 464.6 and generally faster than larger models. Medical fine-tuning is not a guarantee of quality: the Bio-Medical family scored well below the untuned Llama 3.1 8B model, and the reasoning-heavy DeepSeek R1 models were both the slowest and less accurate than their non-reasoning counterparts. On older devices, throughput remains usable for interactive conversation, and where a model fails to run the cause is iOS's per-app memory ceiling rather than chip speed—for example, a model needing 2.24 GB was terminated on a 4 GB iPhone with a 2.25 GB app limit.

Load-bearing premise

The accuracy ranking rests on the assumption that GPT-4o's scores under the modified testing conditions—temperature 0.1, no follow-up reask, stripped reasoning text, and compressed 4-bit models—measure clinical reasoning fairly, without hurting some models more than others.

Editorial extensions

If this is right

  • Privacy-preserving clinical question answering is already possible on current iPhones and iPads, without sending patient data to a cloud service.
  • A model's memory footprint, not its parameter count or the device's processor, decides whether it can run, so future devices with more RAM will widen the set of usable medical models more than faster chips will.
  • Medical fine-tuning labels do not predict benchmark performance; the untuned Llama 3.1 8B beat several medically fine-tuned models, so real deployments need per-model validation on realistic clinical cases.
  • Throughput of 8 to 30 tokens per second on everyday devices exceeds average reading speed, making interactive on-device use practical in clinical settings.
  • Reasoning-style models such as DeepSeek R1 are slower and not more accurate on this benchmark, suggesting that long chain-of-thought outputs are a poor use of limited edge compute.

Reading between the lines

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

  • If the modified GPT-4o scoring is fair, the best on-device 8B models score about 500 AMEGA points, roughly 70 percent of the score reported for a much larger cloud model under the original protocol—enough to suggest triage-level support, though not equivalence.
  • Removing the reask step may systematically favor smaller models, since reasking prolongs outputs and can trigger hallucinations; re-running with reask enabled could change the ranking even if the feasibility result stands.
  • The thermal measurements—throughput drops of 18 to 55 percent in the serious thermal state—imply that real deployments need thermal-aware scheduling or duty cycling; testing whether short bursts preserve both speed and accuracy would be a natural extension.
  • Because all devices are from one vendor's ecosystem, the memory-limit conclusion may not transfer to other operating systems with different memory management; repeating the same 13-model benchmark on another mobile runtime would test that boundary.
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

4 major / 6 minor

Summary. The manuscript benchmarks 13 open-weight LLMs running on-device via Apple's MLX framework across six iOS devices, scoring clinical-reasoning performance with the AMEGA benchmark. The authors report that medical models Med42 and Aloe achieve the highest AMEGA scores, that Phi-3 Mini offers a strong accuracy/speed trade-off, and that memory, rather than raw compute, is the main deployment constraint. The paper also logs throughput, thermal state, and memory behavior across devices.

Significance. If the accuracy ranking were validated, this paper would provide a practically useful map for deploying clinical LLMs on consumer Apple hardware, with genuine value in the open-source HealthBench toolchain and the realistic real-device measurements of throughput, thermal state, and memory limits. The feasibility findings—older devices can run smaller models, memory is the binding constraint—rest on direct measurements and are reasonably solid. The accuracy ranking, however, rests on an unvalidated modification of the AMEGA protocol and on single-run scores without variance reporting, so the headline model-comparison claims are not yet trustworthy. For a clinical AI venue, the benchmark rigor must be strengthened before the accuracy claims can be accepted.

major comments (4)
  1. [§2.2, §2.5, Table 3] The central accuracy ranking is based on AMEGA scores collected under a protocol that differs from the published benchmark in four consequential ways: generation temperature is 0.1 instead of 0.0 (§2.2), the reask process is removed (§2.1), <think> blocks are stripped from reasoning models (§2.5), and all converted models are quantized to 4-bit (§2.3). The authors themselves state that temperature 'can significantly influence output quality' (§4.3) and that the reask exclusion 'require[s] further validation' (§2.1). No control sub-study is reported that quantifies the effect of these deviations on the scores. Consequently, the reported ordering—for instance, the 25.9-point gap between Med42/Aloe and Phi-3 Mini—cannot be attributed to clinical reasoning ability rather than to interactions between the modified protocol and model family. This is the load-bearing assumption for the abstract and §4.1 claims.
  2. [§3.1, Table 3] The accuracy scores are reported only as means across devices, with no standard deviation, confidence interval, or per-device breakdown. Since each model-device combination appears to have been run once and the generation temperature is 0.1, sampling variability is entirely uncharacterized; the ordering of models with scores within roughly 25 points may be within noise. Additionally, models do not run on the same set of devices (e.g., Med42 runs only on devices with at least 8 GB, while Phi-3 Mini also runs on smaller devices), so the means are not directly comparable. The authors should report per-device scores and variability, ideally with repeated runs or a statistical comparison on a matched device subset.
  3. [§2.3, Table 1] The quantization is applied unevenly: DeepSeek R1 1.5B is kept at 8-bit while all other converted models use 4-bit. This breaks comparability of that model's score (179.8) with Llama 3.2 1B (256.5), and no ablation is provided to measure the accuracy impact of 4-bit quantization for any model. Section 4.3 acknowledges that the conversion 'likely led to a reduction in model performance,' but the direction and magnitude of that effect are not quantified. Without such an ablation, the cross-model accuracy comparisons in Table 3 cannot be separated from quantization-induced artifacts.
  4. [§2.5, §4.3] The GPT-4o evaluator is used without a human-verification substudy. Although the original AMEGA benchmark relies on the same judge, the responses judged here come from quantized, on-device models with modified generation parameters, and the judge's behavior on this distribution is uncalibrated. The paper's limitation section concedes that a human evaluation 'could have identified additional issues or resulted in different answer ratings' (§4.3). A small human-rated sample of responses would provide a direct check on whether the judge biases particular model families, which is especially important given the removal of the reask process from the original evaluation pipeline.
minor comments (6)
  1. [§1 vs. §2.1] The Introduction states that AMEGA contains '20 clinical cases across 13 medical specialties,' while §2.1 says the cases cover '20 medical specialties'; these numbers should be reconciled.
  2. [§3.1 vs. Table 3] The model is referred to as 'MedLlama3-v20' in the text but as 'MedLlama 8B' in Table 3; a single consistent identifier should be used throughout.
  3. [§2.2] The sentence beginning 'A cool-down phase is initiated before inference for each questin begins' contains a typo ('questin' for 'question') and should be rewritten for clarity.
  4. [Table 1 vs. Table 2] Table 1 reports 'Memory Usage (GB)' while Table 2 uses 'Memory (GiB)' and 'Memory Limit (GiB)'; the units should be made consistent and explicitly defined.
  5. [Figure 1b] The caption for Figure 1b states that the minimum value 'reflects the average T/s on the slowest device'; it should clarify that this is the mean over questions for that device, not the minimum instantaneous throughput.
  6. [References] Reference [19] spells 'Microsoft' as 'Mircosoft'; other minor citation formatting issues (e.g., inconsistent URL line breaks in [1]) could be cleaned up.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: accuracy claims are measured against an external benchmark (AMEGA) with an external judge (GPT-4o), and no fitted parameter or self-citation chain forces the reported ranking.

full rationale

The paper's central claims are empirical benchmarking results rather than derived predictions. Model scores in Table 3 are obtained by running publicly available, pre-trained models on mobile devices and having GPT-4o evaluate the generated responses with the published AMEGA rubric. No parameter is fit to the reported AMEGA scores, and no model is trained or tuned in this paper. The protocol modifications (temperature raised to 0.1, reask removed, think blocks stripped, 4-bit MLX quantization) are genuine validity threats, and the authors candidly acknowledge this in Section 4.3: 'These parameter modifications can significantly influence output quality, potentially affecting both accuracy and inference speed.' However, a validity threat is not circularity: the reported measurements are not equivalent to the evaluation inputs by construction. There is also no load-bearing self-citation. The HealthBench application and Stanford Spezi LLM module are described as open-source implementation artifacts used to collect the measurements, not as authority for the accuracy findings, and the benchmark itself is attributed to Fast et al. (an external group). The memorization concern in Section 4.1 is a contamination hypothesis, not a circular derivation, and does not make the benchmark scores an input to themselves. Overall, the accuracy, speed, and feasibility conclusions are self-contained empirical observations; the paper's main risks are validity and generalization, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests primarily on external benchmark validity (AMEGA plus GPT-4o judge) and on protocol choices (temperature, quantization, token limits) that are set by hand rather than fitted. No free parameters are fit to the AMEGA outcomes; the listed free parameters are protocol settings that could bias model comparisons.

free parameters (3)
  • Inference temperature = 0.1
    Chosen after pilot testing because temperature 0.0 produced degenerate outputs; deviates from the AMEGA protocol (Sections 2.2 and 4.3) and may affect both accuracy and token generation speed across models.
  • Output token limit = 2048 standard, 4096 reasoning
    Set to cap generation length, mitigate repetitive outputs, and bound memory use (Section 2.2); differences in truncation could affect AMEGA scores for verbose models.
  • MLX quantization bits = 4-bit for all models except DeepSeek R1 1.5B at 8-bit
    MLX default conversion settings (group size 64, float64) chosen for device compatibility (Section 2.3); differing quantization levels likely degrade models unevenly, confounding cross-model comparisons.
assumptions (4)
  • domain assumption AMEGA benchmark scores measure clinical reasoning quality
    The study treats AMEGA scores as ground-truth clinical reasoning quality and uses them to rank models; this assumption is inherited from Fast et al. and not validated for on-device outputs (Section 2.1).
  • domain assumption GPT-4o evaluates model answers without bias
    All responses are graded by GPT-4o at temperature 0 with no human verification; evaluator biases or sensitivity to formatting could systematically favor certain models (Section 2.5).
  • domain assumption MLX framework and unified memory fairly represent on-device LLM performance
    Performance and memory measurements are specific to Apple's MLX runtime; conclusions about feasibility and memory limits may not transfer to other on-device runtimes (Section 2.4).
  • domain assumption Single run per model-device pair is representative
    The paper does not state how many times each model-device pair was run; if only once, the means in Table 3 lack sampling error context (Section 3.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Medicine on the Edge: Comparative Performance Analysis of On-Device LLMs for Clinical Reasoning." pith.science (2026). https://pith.science/paper/XVOKXQKO

@misc{pith2026250208954,
  author       = {Pith},
  title        = {Pith review of: Medicine on the Edge: Comparative Performance Analysis of On-Device LLMs for Clinical Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XVOKXQKO}},
  note         = {Machine review of arXiv:2502.08954}
}
read the original abstract

The deployment of Large Language Models (LLM) on mobile devices offers significant potential for medical applications, enhancing privacy, security, and cost-efficiency by eliminating reliance on cloud-based services and keeping sensitive health data local. However, the performance and accuracy of on-device LLMs in real-world medical contexts remain underexplored. In this study, we benchmark publicly available on-device LLMs using the AMEGA dataset, evaluating accuracy, computational efficiency, and thermal limitation across various mobile devices. Our results indicate that compact general-purpose models like Phi-3 Mini achieve a strong balance between speed and accuracy, while medically fine-tuned models such as Med42 and Aloe attain the highest accuracy. Notably, deploying LLMs on older devices remains feasible, with memory constraints posing a greater challenge than raw processing power. Our study underscores the potential of on-device LLMs for healthcare while emphasizing the need for more efficient inference and models tailored to real-world clinical reasoning.

Figures

Figures reproduced from arXiv: 2502.08954 by the authors.

Figure 1
Figure 1. Performance Evaluation Across Devices and Models. [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Output tokens per second compared to the thermal state of all iPhones over all [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison of LLMs across different devices. Plot ( [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Preserving Privacy, Increasing Accessibility, and Reducing Cost: An On-Device Artificial Intelligence Model for Medical Transcription and Note Generation

    cs.CL 2025-07 conditional novelty 4.0 of 10

    Fine-tuning a 1B Llama model on synthetic endocrinology data improves structured medical note generation and substantially reduces LLM-judged hallucinations and omissions in a browser-based, on-device deployment.

Reference graph

Works this paper leans on

30 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    2023, year of open LLMs, December 2023

    Clémentine Fourrier. 2023, year of open LLMs, December 2023. URL https:// huggingface.co/blog/2023-in-llms

    work page 2023

  2. [2]

    An Empirical Analysis and Resource Footprint Study of Deploying Large Language Models on Edge Devices

    Nobel Dhar, Bobin Deng, Dan Lo, Xiaofeng Wu, Liang Zhao, and Kun Suo. An Empirical Analysis and Resource Footprint Study of Deploying Large Language Models on Edge Devices. In Proceedings of the 2024 ACM Southeast Conference, ACMSE ’24, pages 69–76, New York, NY, USA, April 2024. Association for Com- puting Machinery. ISBN 9798400702372. doi: 10.1145/3603...

    work page arXiv 2024

  3. [3]

    PrivateLoRA For Effi- cient Privacy Preserving LLM, November 2023

    Yiming Wang, Yu Lin, Xiaodong Zeng, and Guannan Zhang. PrivateLoRA For Effi- cient Privacy Preserving LLM, November 2023. URLhttp://arxiv.org/abs/2311. 14030. arXiv:2311.14030 [cs]

    arXiv 2023

  4. [4]

    Gostin, Sam F

    Lawrence O. Gostin, Sam F. Halabi, and Kumanan Wilson. Health Data and Privacy in the Digital Era.JAMA, 320(3):233–234, July 2018. ISSN 0098-7484. doi: 10.1001/ jama.2018.8374. URL https://doi.org/10.1001/jama.2018.8374

    work page arXiv 2018

  5. [5]

    How Much Energy Do LLMs Consume? Unveil- ing the Power Behind AI, July 2024

    Sourabh Mehta. How Much Energy Do LLMs Consume? Unveil- ing the Power Behind AI, July 2024. URL https://adasci.org/ how-much-energy-do-llms-consume-unveiling-the-power-behind-ai/

    work page 2024

  6. [6]

    It’s time to bench the medical exam benchmark

    Inioluwa Deborah Raji, Roxana Daneshjou, and Emily Alsentzer. It’s time to bench the medical exam benchmark. NEJM AI, 2(2):AIe2401235, January 2025. doi: 10.1056/AIe2401235. URL https://ai.nejm.org/doi/full/10.1056/AIe2401235. Publisher: Massachusetts Medical Society

    doi:10.1056/aie2401235 2025

  7. [7]

    Desai, Kueiyu Joshua Lin, and Jie Yang

    Bowen Gu, Rishi J. Desai, Kueiyu Joshua Lin, and Jie Yang. Probabilistic medical predictions of large language models.npj Digital Medicine, 7(1):1–9, December 2024. ISSN 2398-6352. doi: 10.1038/s41746-024-01366-4. URLhttps://www.nature.com/ articles/s41746-024-01366-4. Publisher: Nature Publishing Group

    work page doi:10.1038/s41746-024-01366-4 2024

  8. [8]

    Adams, Felix Busch, Conor Fallon, Marc Huppertz, Robert Siepmann, Philipp Prucker, Nadine Bayerl, Daniel Truhn, Marcus Makowski, 16 Alexander Löser, and Keno K

    Dennis Fast, Lisa C. Adams, Felix Busch, Conor Fallon, Marc Huppertz, Robert Siepmann, Philipp Prucker, Nadine Bayerl, Daniel Truhn, Marcus Makowski, 16 Alexander Löser, and Keno K. Bressem. Autonomous medical evaluation for guideline adherence of large language models. npj Digital Medicine, 7, 2024. doi: 10.1038/s41746-024-01356-6. URL https://www.nature...

    doi:10.1038/s41746-024-01356-6 2024

Show all 30 references
  1. [9]

    Smoothquant: Accurate and efficient post-training quantization for large language models, 2024

    Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and efficient post-training quantization for large language models, 2024. URL https://arxiv.org/abs/2211.10438

    2024 arXiv

  2. [10]

    MLX: Efficient and flexible machine learning on apple silicon, 2023

    Awni Hannun, Jagrit Digani, Angelos Katharopoulos, and Ronan Collobert. MLX: Efficient and flexible machine learning on apple silicon, 2023. URLhttps://github. com/ml-explore

    2023

  3. [11]

    ImportingPhan- toms: Measuring LLM Package Hallucination Vulnerabilities, January 2025

    ArjunKrishna, ErickGalinkin, LeonDerczynski, andJeffreyMartin. ImportingPhan- toms: Measuring LLM Package Hallucination Vulnerabilities, January 2025. URL http://arxiv.org/abs/2501.19012. arXiv:2501.19012 [cs] version: 1

    2025 arXiv

  4. [12]

    Your battery status is being used to track you online.The Guardian, August 2016

    Alex Hern. Your battery status is being used to track you online.The Guardian, August 2016. ISSN 0261-3077. URL https://www.theguardian.com/technology/ 2016/aug/02/battery-status-indicators-tracking-online

    2016

  5. [13]

    Domain-specialized llm: Financial fine-tuning and utilization method using mistral 7b

    Cheonsu Jeong. Domain-specialized llm: Financial fine-tuning and utilization method using mistral 7b. Journal of Intelligence and Information Systems, 30(1):93–120, March 2024. ISSN 2288-4882. doi: 10.13088/jiis.2024.30.1.093. URL http://dx. doi.org/10.13088/jiis.2024.30.1.093

    2024 doi

  6. [14]

    Bio-medical-llama-3-2-1b-cot-012025: A reasoning-enhanced biomed- ical language model, 2025

    ContactDoctor. Bio-medical-llama-3-2-1b-cot-012025: A reasoning-enhanced biomed- ical language model, 2025. URL https://huggingface.co/ContactDoctor/ Bio-Medical-Llama-3-2-1B-CoT-012025

    2025

  7. [15]

    Llama-3.2-1b-instruct, 2024

    Meta Llama. Llama-3.2-1b-instruct, 2024. URL https://huggingface.co/ meta-llama/Llama-3.2-1B-Instruct

    2024

  8. [16]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei F...

    2025 arXiv

  9. [17]

    Bio-medical-3b-cot-012025: A high-performance biomedical lan- guage model with reasoning capabilities, 2025

    ContactDoctor. Bio-medical-3b-cot-012025: A high-performance biomedical lan- guage model with reasoning capabilities, 2025. URL https://huggingface.co/ ContactDoctor/Bio-Medical-3B-CoT-012025

    2025

  10. [18]

    Llama-3.2-3b-instruct, 2024

    Meta Llama. Llama-3.2-3b-instruct, 2024. URL https://huggingface.co/ meta-llama/Llama-3.2-3B-Instruct

    2024

  11. [19]

    Phi-3-mini-4k-instruct, 2024

    Mircosoft. Phi-3-mini-4k-instruct, 2024. URL https://huggingface.co/ microsoft/Phi-3-mini-4k-instruct

    2024

  12. [20]

    Qwen2 technical report.arXiv preprint arXiv:2407.10671, 2024

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng- peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzhe...

    2024 arXiv

  13. [21]

    Bio-medical-llama-3-8b: A high-performance biomedical language model, 2024

    ContactDoctor. Bio-medical-llama-3-8b: A high-performance biomedical language model, 2024. URL https://huggingface.co/ContactDoctor/ Bio-Medical-Llama-3-8B

    2024

  14. [22]

    Aloe: A family of fine-tuned open healthcare llms, 2024

    Ashwin Kumar Gururajan, Enrique Lopez-Cuena, Jordi Bayarri-Planas, Adrian Tor- mos, Daniel Hinjos, Pablo Bernabeu-Perez, Anna Arias-Duart, Pablo Agustin Martin- Torres, Lucia Urcelay-Ganzabal, Marta Gonzalez-Mallo, Sergio Alvarez-Napagao, Ed- uard Ayguadé-Parra, and Ulises Cor...

    2024

  15. [23]

    medllama3-v20, 2024

    YonseiMAILAB. medllama3-v20, 2024. URL https://huggingface.co/ ProbeMedicalYonseiMAILab/medllama3-v20

    2024

  16. [24]

    Med42-v2: A suite of clinical llms, 2024

    Clément Christophe, Praveen K Kanithi, Tathagata Raha, Shadab Khan, and Marco AF Pimentel. Med42-v2: A suite of clinical llms, 2024. URL https: //arxiv.org/abs/2408.06142

    2024 arXiv

  17. [25]

    Llama-3.1-8b-instruct, 2024

    Meta Llama. Llama-3.1-8b-instruct, 2024. URL https://huggingface.co/ meta-llama/Llama-3.1-8B-Instruct

    2024

  18. [26]

    Del Mundo, Mohammad Rastegari, and Mehrdad Farajtabar

    Keivan Alizadeh, Iman Mirzadeh, Dmitry Belenko, Karen Khatamifard, Minsik Cho, Carlo C. Del Mundo, Mohammad Rastegari, and Mehrdad Farajtabar. LLM in a flash: Efficient Large Language Model Inference with Limited Memory, July 2024. URL http://arxiv.org/abs/2312.11514. arXiv:23...

    2024 arXiv

  19. [27]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, January 2023

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, January 2023. URLhttp://arxiv.org/abs/2201.11903. arXiv:2201.11903 [cs]

    2023 arXiv

  20. [28]

    Bouw, and Stephen Cobb

    Tolga Çöplü, Marc Loedi, Arto Bendiken, Mykhailo Makohin, Joshua J. Bouw, and Stephen Cobb. A Performance Evaluation of a Quantized Large Language Model on Various Smartphones, December 2023. URL http://arxiv.org/abs/2312.12472. arXiv:2312.12472 [cs]

    2023 arXiv

  21. [29]

    Hl7 fhir: An agile and restful approach to healthcare information exchange

    Duane Bender and Kamran Sartipi. Hl7 fhir: An agile and restful approach to healthcare information exchange. In Proceedings of the 26th IEEE Interna- tional Symposium on Computer-Based Medical Systems, pages 326–331, 2013. doi: 10.1109/CBMS.2013.6627810

    2013

  22. [30]

    How many words do we read per minute? a review and meta- analysis of reading rate.Journal of Memory and Language, 109, 2019

    Marc Brysbaert. How many words do we read per minute? a review and meta- analysis of reading rate.Journal of Memory and Language, 109, 2019. doi: 10.1016/ j.jml.2019.104047. URL https://www.sciencedirect.com/science/article/abs/ pii/S0749596X19300786#preview-section-cited-by. 19

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.