A generated title card reading 'LingSante-3.1-flash vs LFM3 1.7B Base' with the subtitle 'Someone else's corpus, or yours'. Two rounded panels compare them: LingSante-3.1-flash shows 'Strategy: buy the knowledge', 'Parameters: 560B total, 25B active' and 'Adaptation: serve it, no recipe published'; LFM3 1.7B Base shows 'Strategy: build on your data', 'Parameters: 1.73B' and 'Adaptation: fine-tuning notebooks included'. A footer reads 'Buy costs hardware; build costs data and evaluation.' The OrcaRouter logo is composited in a white strip along the bottom edge.
Guides & Insights

LingSante-3.1-flash vs LFM3 1.7B Base: Buy vs Build a Medical Model

Autor

jinhao.song

Data publikacji

Najnowsze modele · 20Zobacz wszystkie modele →
Benchmarki: Artificial Analysis · aktualizowane codziennie
Powrót do wszystkich wpisów

There are two ways to get a medical model that knows your patients. You can buy the knowledge pre-trained: LingSante-3.1-flash, a 560-billion-parameter medical mixture-of-experts model that inclusionAI uploaded on 19 November 2026, trained on a medical corpus and shipped under MIT. Or you can build it: take LFM3 1.7B Base, Liquid AI's 1.73-billion-parameter pre-trained checkpoint released on 27 October 2026, and fine-tune it on your own notes, guidelines and outcomes. The first buys someone else's corpus. The second buys your own.

That is a strategic choice, not a spec comparison, and it is the reason this pairing is more useful than its parameter gap suggests. A 1.7B base checkpoint and a 560B specialist are not competing for the same job. They represent two different answers to the question every clinical AI team eventually faces: do you need general medical knowledge, or do you need knowledge of this hospital?

The two strategies, priced

• Buy — LingSante-3.1-flash arrives trained. No data collection, no labelling budget, no training run. You pay in hardware: 291.72 GiB in MXFP4 on four H20-3e GPUs, or roughly 578 GB in FP8 on eight H200s. The medical knowledge is whatever Ant Group put in the corpus, and you cannot inspect that corpus from the repository.

• Build — LFM3 1.7B Base arrives empty of medical specialization but complete as a starting point. You pay in data and engineering: a curated corpus, a fine-tuning pipeline, evaluation sets, and the clinical expertise to label them. The compute is small — the base checkpoint is about 3.5 GB in BF16 and roughly 1 GB at four-bit precision, so a single GPU is enough for parameter-efficient tuning.

• The asymmetry — buying is cheap in engineering and expensive in hardware and opacity. Building is cheap in hardware and expensive in data and expertise. Neither is cheaper overall; they front-load different costs.

For a team with ten years of structured records and no GPUs, the build path is the only one that fits the budget. For a team with GPUs and no curated corpus, the buy path is the only one that ships this quarter.

What LFM3 1.7B Base gives you to build on

Liquid AI's card is unusually explicit about what a base checkpoint is for. It states the model is "only recommended for tasks that require heavy fine-tuning" and points chat and agent use at the instruction-tuned sibling.

• Architecture — 24 layers in a tri-hybrid arrangement: 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. The design assumes most of a small model's work can be done without full attention, which is what makes the on-device footprint plausible.

• Training — 22 trillion tokens. The final 3.5 trillion-token mid-training stage added agentic traces and tool calls and extended the context window to 262,144 tokens.

• Interface — a 128,000-token vocabulary across sixteen languages, a chat template shared with the tuned sibling, and a documented switch: enable_thinking=False prefills an empty thinking block rather than leaving the template to guess.

• On-ramp — model code, tokenizer, generation config, and a family of TRL fine-tuning notebooks. The card tells you to change one line — model_id = "LiquidAI/LFM3-1.7B-Base" — to start from the base weights instead of the tuned ones.

• Licence — the LFM Open License v1.0, which is not OSI-open in the way Apache 2.0 is. It has commercial terms worth reading before a product decision.

The important part is the seam. Liquid documents where pre-training ends and post-training begins, so a clinical team can insert its own data at a known point. That is what a base checkpoint is for, and this one comes with instructions.

Screenshot of the Hugging Face model card for LiquidAI/LFM3-1.7B-Base captured 22 November 2026 in English: header with 169 likes, 16 languages, lfm3, liquid and edge tags, License lfm1.0 and 15,475 downloads last month; sidebar showing 2B parameters in BF16 and 'This model isn't deployed by any Inference Provider'; body text describing a pre-trained checkpoint trained on 22T tokens with the final 3.5T mid-training tokens extending context to 256K, about 1 GB at 4-bit, and a Model details table listing 1.73B parameters, 24 layers (13 gated short convolutions, 5 sliding-window attention, 6 global attention), 256K context, 128,000 vocabulary, sixteen languages and the LFM Open License v1.0.

What LingSante-3.1-flash gives you to skip

The inclusionAI model skips the entire build phase. Its card describes medical and general data mixed through pretraining and mid-training, followed by supervised fine-tuning, medical reinforcement learning, and multi-teacher on-policy distillation. In principle that is years of corpus work you do not have to repeat.

• Published capability — twenty vendor-run benchmark rows. HealthBench Professional 65.72 is the highest in its comparison row; MedQA-USMLE 94.50; MedXpertQA-Text 57.27; BrowseComp with context 94.10. The same table shows weaker agentic and coding results: Terminal-Bench 4.0 at 26.5, SWE-Bench Pro at 65.25, GDPval-AA v2.1 at 1,600 Elo.

• What it knows — a general medical corpus. Guidelines, exam-style knowledge, professional communication, tool-using workflows. What it does not know is your formulary, your discharge template, your local referral pathways, or the idiosyncrasies of your patient population.

• What it costs to adapt — the card describes no fine-tuning recipe, no LoRA configuration, and no adapter support. The repository ships inference code, not a training pipeline. You can serve it; adapting it is an open engineering question.

• The evidence caveat — every number is vendor-reported and unreproduced. There is no independent evaluation and, as of 22 November, zero downloads recorded on the repository.

The "buy" option is therefore less turnkey than it sounds. You skip pre-training, but you still own evaluation, integration, and the gap between general medical knowledge and local practice. What you genuinely buy is a starting point that already reasons about medicine.

Screenshot of the Hugging Face model card for inclusionAI/LingSante-3.1-flash captured 22 November 2026 in English: 572B parameters in the sidebar, License mit, 'This model isn't deployed by any Inference Provider', and an introduction stating 560 billion total parameters, 25 billion per token and a 1-million-token context window, followed by the research-and-evaluation-only notice and the embedded medical benchmark chart.

The data question neither solves

Both paths converge on the same unresolved problem: you need a labelled evaluation set from your own setting, and neither vendor supplies one.

For the build path, the evaluation set is also the training set's quality control. A 1.7B fine-tune on 5,000 discharge summaries will learn the template and the local vocabulary. Whether it learns the clinical judgment is a separate question, and answering it requires held-out cases with expert labels. That is the expensive part of "build", and it does not shrink because the model is small.

For the buy path, the evaluation set is the only defence against a vendor claim. LingSante-3.1-flash scoring 65.72 on HealthBench Professional says nothing about your triage notes. You need your own cases, your own reviewers, and a threshold for what counts as acceptable. The 292 GB deployment is the easy part of that budget by comparison.

There is also a privacy asymmetry. The build path keeps every training example inside your perimeter, provided you fine-tune on your own hardware. The buy path keeps inference inside your perimeter if you self-host, but the model's knowledge came from a corpus you cannot inspect. For a hospital, the first is usually easier to explain to a data-protection officer; the second is usually easier to explain to a clinician who wants the model to know medicine on day one.

A decision that depends on volume

The economics turn on how many distinct tasks you have.

If you have one narrow, high-volume task — coding referrals, extracting lab values, drafting standard letters — a fine-tuned 1.7B model is likely to win. It runs cheaply, can sit on-premises, and a narrow task is exactly what a small fine-tune does well. The 560B specialist is overkill, and its hardware floor makes it uneconomic for a single workflow.

If you have many tasks across specialties, the calculus flips. Fine-tuning a separate small model for each one multiplies the data and evaluation work. A general medical model that handles reasoning, tool use and long records out of the box amortises its deployment cost across the portfolio. That is the case LingSante-3.1-flash is built for, and the case where 292 GB starts to look reasonable.

If you have neither — no corpus and no GPUs — the honest answer is that neither path is available yet, and the first purchase should be data infrastructure rather than a model.

How to pilot both without a rebuild

The two paths are not mutually exclusive, and the sensible pilot uses each where it is cheap.

Run the build path first on a single GPU with a narrow task and a held-out evaluation set. That tells you whether your data is good enough to specialise a small model, which is a question no vendor can answer for you. If the fine-tune clears your threshold, you have a deployable, on-premises component and a data pipeline you will need anyway.

Then evaluate the buy path on the tasks the small model cannot cover. That requires the multi-GPU deployment, so treat it as a capital decision rather than a trial. OrcaRouter does not host LingSante-3.1-flash, so there is no shortcut through us for that half. Where a gateway does help is the surrounding production path: if your application also calls hosted general models, one endpoint for 200-plus models with automatic failover keeps a self-hosted fine-tune from becoming a single point of failure, and 0% markup means the provider's list price is what you pay. The routing layer is the plumbing around the decision, not a substitute for it.

The choice, stated plainly: build if your advantage is proprietary data and your task is narrow; buy if your need is broad medical reasoning and your constraint is time rather than hardware. Most clinical organisations will eventually do both, and the ones that succeed will be the ones that built the evaluation set before they chose the model.

A generated two-column scoreboard titled 'LingSante-3.1-flash vs LFM3 1.7B Base - the scoreboard'. Six shared rows compare 'Strategy: buy the knowledge' against 'Strategy: build on your data'; 'Parameters: 560B total, 25B active' against '1.73B'; 'Starting point: medical corpus' against 'pre-trained only'; 'Adaptation: no training recipe' against 'TRL notebooks included'; 'Hardware: 4 H20 GPUs' against 'one GPU'; and 'Licence: MIT' against 'LFM Open License v1.0'. A footer reads 'LingSante figures vendor-reported; LFM3 Base has no medical evaluation.' The OrcaRouter logo is composited in a white strip along the bottom edge.