
LingSante-Pulse-1.0 Is on Hugging Face; the Code to Run It Isn't
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 183 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 741 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 407 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 150 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 83 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 119 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 108 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 212 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
LingSante-Pulse-1.0 appeared on Hugging Face at 08:39 UTC on 20 November 2026, and by 09:12 the same morning its author, inclusionAI, had already pushed a revision. The listing reports one like and zero downloads. Nothing in the repository is a model you can talk to: LingSante-Pulse-1.0 is a 16.7-million-parameter encoder for physiological signals — ECG, EEG, airflow, blood oxygen — that turns a waveform into a fixed-length embedding and stops there. There was no launch post, no technical report, no benchmark table, and, as of this writing, no code package to load it with. The card instructs you to install a separate lingsante_pulse package; that package is not on PyPI.
What does exist is unusually legible. The weights ship in F32, the safetensors header is uncompressed and readable, and channels.json names all fifty-seven signal types the model accepts. So it is possible to say quite a lot about what this model is, and to be precise about what nobody outside the lab can yet confirm. This piece keeps those two categories separate.
What appeared, and what did not
The repository holds seven files: a README, an MIT licence, model.safetensors (63.8 MiB of weights in a 66.9 MB repository), config.json, channels.npz, channels.json, and the usual gitattributes stub. That is the whole release.
Absent, and each absence is load-bearing:
• No inference code in the repository, and no package on PyPI under either lingsante-pulse or lingsante_pulse — both return 404.
• No benchmarks. The card describes behaviour in prose ("turns physiological signals such as ECG into embeddings") and publishes no numbers of any kind.
• No technical report. inclusionAI has an arXiv presence, and its sibling LingSante-MedGuard repository, created on 17 November 2026, links a report. The Pulse card links nothing.
• No announcement of any kind that a web search can find. The model card itself contains the phrase "After publication", which reads like text written before the package's release was scheduled.
• No inference provider. Hugging Face reports the model as not deployed by any provider, and the card invites readers to request support.
One file is a small surprise: config.json carries SHA-256 digests for the three data files. That is not a marketing artefact. It is the kind of check a team adds when it expects the files to be copied around and wants to know they arrived intact.
Twelve blocks, 256 dimensions, 32-sample patches
The safetensors header lists 203 tensors and its metadata block reads model: LingSante-Pulse, step: 260000. Reconstructing the architecture from the tensor names is straightforward — and it is the only architecture documentation that exists.
• Depth and width — 12 transformer blocks (blocks.0 through blocks.11), each with a 256-wide residual stream.
• Attention — a fused qkv projection of shape [768, 256], so three 256-wide heads' worth of keys, queries and values; no separate output projection size is given beyond out at [256, 256].
• Feed-forward — gate_up at [2048, 256] and down at [256, 1024], the shape signature of a gated linear unit with a 1,024-wide hidden layer.
• Gating — every block carries an attention_gate at [256, 256] plus bias, and a matching film_heads.N at [1024, 256] emitting a 1,024-value modulation vector. Twelve FiLM heads for twelve blocks. Nothing in the card explains what conditions them.
• Patching — patch.weight is [256, 32]. Thirty-two samples per patch, at the fixed 128 Hz the card specifies, is exactly 250 milliseconds, which is the interval the README gives for returned token features. The arithmetic closes.
• Sequence furniture — one cls token, one mask_token, and four register_tokens in a [1, 4, 256] tensor.
• A text path — text_adapter.0 projects 384 dimensions into 256, and text_adapter.1 refines within 256. The card says the channel-condition vectors come from BAAI/bge-small-en-v1.5, whose hidden size is 384. So the model is conditioned on natural-language descriptions of its input channels, projected into its own space. That reading is inference from shapes plus the card's provenance note, not a documented design, but the two agree.
• One more head — validity.weight at [256, 64]. A per-position signal that a sample was actually observed, which matches the card's talk of observed-valid PatchMean and storage-padding masks.

A spec sheet read off tensor shapes rather than a datasheet: 16.7M parameters, twelve blocks at width 256, 32-sample patches at 128 Hz, and a 384-to-256 text adapter for channel conditioning.
Fifty-seven channels describe a sleep lab
The interesting file is not the weights. It is channels.json, which carries a plain-English description of each of the 57 signal identities the model accepts.
Read the list in order and it stops looking like a generic biosignal vocabulary. Three EEG scalp leads "during sleep". Two EOG leads "during sleep". A chin EMG "during sleep", a leg EMG "during sleep". Airflow, thoracic effort, abdominal effort, snore, SpO2. That is a polysomnography montage, in the order a sleep technician would name it.
Then the list keeps going into a different room. Airway pressure "measured by an anesthesia ventilator". End-tidal CO2. Respiratory rate derived from a Dräger Primus and from a Solar8000 monitor. BIS EEG power in decibels from a forehead depth-of-anaesthesia monitor, plus its two frontal EEG leads. Twelve-lead ECG with a full precordial set, three modified limb leads, and sixteen monopolar EOG derivations.
Two settings, one channel vocabulary. The most defensible inference available from this repository is that the encoder was trained across both sleep studies and perioperative or intensive-care monitoring — and that its designers wanted one model that survives being pointed at whichever montage a hospital already owns. The card does not say this. The channel list says it for them.
That breadth is also the model's main engineering claim. A 16-million-parameter encoder that accepts 57 heterogeneous channel identities, resamples everything to 128 Hz, and returns one embedding per channel is solving a data-plumbing problem that anyone who has worked with clinical waveforms recognises immediately: nothing in a hospital agrees on sampling rate, lead naming, or which channels were actually recorded.
The missing piece is the package
The card's usage example is four lines long:
from lingsante_pulse import LingSantePulse, then LingSantePulse.from_pretrained(), then model.encode(x, sample_rate=128, channels=[{"modality": "ECG", "lead": "II"}]), then print the shapes.
None of that runs today. The import fails because the package does not exist on the index the card implicitly points at. The model directory contains weights and conditioning vectors, not a Python module. A reader who wants to use this release has exactly one route: reimplement the forward pass from the 203 tensor names, guess the activation functions, guess how the FiLM modulation is applied and in what order, and hope the normalisation statistics are the ones the authors used.
For a transformer over patches that is not impossible. It is a day of work for someone fluent in PyTorch, and a research project for anyone else. But "possible to reimplement" and "shipped" are different things, and the gap between them is where most of the value of a release lives. Weights without an inference path are a claim about a model, not a model you can call.
The card is aware of the seam. "Install the separate inference-code package before using these model assets," it says, and then, in the example, "After publication, this complete example downloads the model on first use." Both sentences point forward. Either the package is coming, or the release is a pre-positioning move — the weights placed in public so they are citable, measurable by an internal harness, and ready the moment the code lands.
What the card claims, and what it omits
The documented interface is narrow and specific, which is a point in its favour.
• Input forms — [T], [C, T], [B, C, T], or aligned channel dictionaries. Ragged recordings work; you tell the model the true sample rate and the physical identity of each channel.
• Preprocessing — the code package is said to handle resampling to a fixed 128 Hz, observed samples, and storage-padding masks. Note that this is a claim about code that does not exist yet.
• Outputs — CLS and observed-valid PatchMean returned separately as [B, C*256], with channels concatenated in input order; return_tokens=True also yields patch features and validity masks at 0.25-second spacing.
• Scope — "These are signal representations; task-specific prediction heads are not included." In plain terms: Pulse gives you features, and you supply the classifier, the detector, the risk score.
What it does not give you: any indication of what those features are good for. There is no linear-probe table, no downstream task list, no statement of which dataset the encoder was trained on, no window length guidance beyond the 1,280-sample example (ten seconds at 128 Hz), no statement about how it handles channels it was never trained on, and no comparison to any existing physiological foundation model. The step counter says 260,000 training steps. That is the only number about training in the entire release.
Why a weights-only drop is hard to score
An encoder is not evaluable in the abstract. Its quality is entirely a function of the downstream task you attach to it, and the standard way to report that is a frozen-feature linear probe across a handful of public datasets with a stated split. Without those numbers, three things are unknowable.
First, whether the embeddings carry information or merely variance. A 256-dimensional vector extracted from ten seconds of ECG can look plausible — smooth, structured, stable across patients — and still encode nothing but heart rate and signal quality. Only a probe distinguishes those.
Second, whether the channel-conditioning helps or is decorative. The architecture invests in a text adapter and 57 channel identities. If a single-channel ECG-only baseline matches the full model, that machinery is cost without benefit, and only an ablation shows it.
Third, whether the model generalises past its training distribution. Clinical waveforms shift hard between devices, electrode placements and populations. "Works on the montage it saw" is a much weaker claim than "works on the montage in your cupboard", and nothing here tests the second.
None of this is an accusation. A model uploaded hours ago with no paper is a normal first step, and the technical report may already be in review. It is simply the honest state of the evidence, and a reader deciding whether to build on this should know that the only thing currently verified is the file listing.
What would change this
Four things, in ascending order of how much they would tell you.
A PyPI release under lingsante_pulse would make the model runnable and turn the card's usage example from documentation into a test. That is the smallest and most probable next step.
A paragraph in the card naming the training corpora would let a reader judge domain match before spending a day on integration. Right now the channel list implies sleep and anaesthesia data without confirming it.
A frozen-feature probe on two or three public datasets — one sleep, one cardiac — would convert the release from an artefact into a result. This is the item that matters most and costs the authors least.
An independent reproduction would matter more than any of the above, and cannot be arranged by the vendor. Nobody has run this model in public. That is the sentence to keep in mind for the next few weeks.
Where does an encoder like this sit in a working system? It produces the features; something else has to reason over them, and in most clinical pipelines that something is a language model reading a summary of the embedding alongside the patient's notes. That reasoning layer is the part OrcaRouter serves — one API across 200-plus models, provider list price passed through with no markup added, and automatic failover when a provider degrades. LingSante-Pulse-1.0 itself is self-hosted weights and is not in our catalogue; the encoder and the model that reads its output are separate purchasing decisions, and only the second one has an API at all.
For now the practical answer is unglamorous. If you have PSG or intraoperative waveform data and PyTorch fluency, download the safetensors file and read the header — it is the most informative artefact in the release. If you need something you can call this week, there is nothing here yet. The 63.8 MiB of weights is a promise with a file size.

The complete release as it stands: seven files, one of them weights, none of them code.

What the repository supports today: a parameter count, an architecture read off tensor shapes, and a channel vocabulary — but no benchmark, no package and no provider.
