
LingSante-3.1-flash vs LingSante Pulse 1.0: One Family, Two Halves
- OrcaNEUOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 pro 1 Mio. Tokens · 186 tok/s
- openaiNEUOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 pro 1 Mio. Tokens · 87 tok/s
- openaiNEUOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 pro 1 Mio. Tokens · 151 tok/s
- tencentNEUTencent: Hy4 Flash2026-11-16$0.02 / $0.07 pro 1 Mio. Tokens · 190 tok/s
- AlibabaNEUQwen: Qwen4 Max2026-11-16$1.50 / $4.50 pro 1 Mio. Tokens · 86 tok/s
- GoogleNEUGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 pro 1 Mio. Tokens · 705 tok/s
- AnthropicNEUAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 pro 1 Mio. Tokens · 246 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 pro 1 Mio. Tokens · 97 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 pro 1 Mio. Tokens · 138 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 pro 1 Mio. Tokens · 85 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 pro 1 Mio. Tokens · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 pro 1 Mio. Tokens · 102 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 pro 1 Mio. Tokens · 70 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 pro 1 Mio. Tokens · 35 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 pro 1 Mio. Tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 pro 1 Mio. Tokens · 38 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligenz
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 pro 1 Mio. Tokens · 231 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligenz
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 pro 1 Mio. Tokens · 101 tok/s
Two repositories, twenty-eight hours apart, from the same lab. On 19 November 2026, inclusionAI uploaded LingSante-3.1-flash, a 560-billion-parameter text model trained for medical reasoning. On 20 November, it uploaded LingSante-Pulse-1.0, a 16.7-million-parameter encoder that turns physiological waveforms into embeddings. The parameter gap is about 33,000 to one. That number invites a comparison, and the comparison is the wrong frame: these two artefacts are not rivals. They are the two halves of a clinical pipeline that the lab has not yet published.
One reads words and produces judgments. The other reads electrical signals and produces vectors. Neither can do the other's job, and neither is complete on its own — the text model cannot see an ECG, and the signal encoder ships without the code needed to run it. What makes the pairing worth writing about is that they appeared close enough together to suggest an intended composition, even though the vendor has described no such composition and announced neither release.
One family, two artefacts
• Size — LingSante-3.1-flash is 560B total with 25B active per token. LingSante-Pulse-1.0 is 16.7M parameters, all in F32, in a 63.8 MiB checkpoint.
• Input — flash takes text tokens. Pulse takes sampled waveforms shaped [T], [C,T] or [B,C,T], or a dictionary of aligned channels, each tagged with a physical identity such as an ECG lead or a sleep EEG derivation.
• Output — flash emits tokens: prose, tool calls, reasoning traces. Pulse emits a CLS vector and a PatchMean vector, 256 wide per channel, plus optional patch features on a quarter-second grid at its fixed 128 Hz sampling rate.
• Licence — MIT for both.
• Release state — flash ships a full serving path: a day-zero SGLang image, a vLLM recipe, and an MXFP4 quantized sibling published on 22 November. Pulse ships weights, a config, two channel-conditioning files and a README, with no inference package on PyPI and no benchmark table.
The family name is doing real work here. "LingSante" marks both as medical artefacts from the same group, and the naming implies a shared programme: one model for signal understanding, one for clinical reasoning. But a shared name is not an integration. Nothing in either repository documents a handoff between them, and no joint evaluation exists.
Pulse turns signals into numbers
The encoder is small enough to read, and its specification is unusually legible.
• Architecture — twelve transformer blocks at width 256, a fused query-key-value projection at [768, 256], a gated feed-forward with a 1,024-wide hidden layer, an attention gate per block, a 32-sample patch embedding, a CLS token, a mask token and four register tokens.
• Signal handling — a fixed 128 Hz preprocessing rate. Thirty-two samples per patch gives one embedding every quarter second. A padded batch carries a validity mask, so the model knows which samples were recorded and which were padding.
• Channel identity — 57 named signal identities in channels.json, spanning sleep polysomnography (EEG leads, EOG, chin and leg EMG, airflow, effort, SpO2) and perioperative monitoring (twelve-lead ECG, arterial pressure, EtCO2, BIS depth of anaesthesia, airway pressure).
• Conditioning — twelve FiLM heads emit a modulation vector each, driven by 384-dimensional text embeddings of the channel descriptions from BAAI/bge-small-en-v1.5. The model is conditioned on written descriptions of what each input channel is.
• What is missing — the usage example imports a lingsante_pulse package that returns 404 on PyPI, and the repository contains no Python. There are no published scores, no training-corpus description, and no downstream heads. The only training artefact is a step: 260000 string in the safetensors metadata.
This is a representation model, not a predictor. Its output is useful only if something downstream consumes it — a classifier for arrhythmia, a sleep-staging head, or a projection into a language model's input space.

Flash turns records into decisions
The text model is the opposite: complete enough to serve, opaque enough to be hard to trust.
• Capability claim — medical reasoning, professional healthcare communication, evidence-based research and long-horizon clinical workflows, trained on a mix of medical and general data through pretraining, mid-training, supervised fine-tuning and reinforcement learning.
• Published scores — twenty vendor-run rows. HealthBench Professional 65.72, MedQA-USMLE 94.50, MedXpertQA-Text 57.27, PatientAgentBench 4.722 on a five-point scale, BrowseComp with context 94.10. The same table shows Terminal-Bench 4.0 at 26.5 and GDPval-AA v2.1 at 1,600 Elo, so it is not a general agent model.
• Serving — eight H200s at TP8/EP8 for the FP8 release, four H20-3e for MXFP4. The shipped config sets a 262,144-token window; the card's one-million-token claim needs a serving recipe the repository does not include.
• Evidence status — every score is the vendor's own, unreproduced. The repository has zero recorded downloads and no announcement.
Where Pulse is inspectable and unrunnable, flash is runnable and unverified. That inversion is the most interesting thing about the pair.

Where they would meet in a stack
A plausible clinical pipeline using both would look like this. A monitor or wearable produces a multi-channel recording: say, a twelve-lead ECG plus SpO2 and respiratory effort. Pulse resamples it to 128 Hz, patches it into quarter-second segments, conditions on channel identity, and returns a 256-wide vector per channel per patch. A downstream head — not shipped — converts those vectors into events or findings. Those findings become text. The text, together with the patient's record and the relevant guideline, goes to LingSante-3.1-flash, which reasons over the combined record, calls a search tool if needed, and drafts a note or a recommendation for human review.
That architecture is coherent. It also puts each model where its weakness is least harmful. Pulse never has to reason; flash never has to parse a waveform. The hard parts are the two seams: the head that turns embeddings into clinical findings, and the projection that turns findings into faithful text. Neither seam is documented by the vendor, and the first one does not exist in either repository.
There is a cheaper alternative for teams that already have a multimodal general model: render the waveform as an image and let a vision model read it. That works for rhythm strips and loses sample-level morphology in the process — amplitude quantised to pixel rows, time to pixel columns. Pulse exists to avoid that conversion. Whether avoiding it produces better clinical output is exactly the experiment nobody has run.
The missing pieces on both sides
The two releases fail in opposite directions, which makes a joint deployment harder than either card suggests.
Pulse is missing execution. Without a published package, the weights cannot be loaded by any documented code path. A team could reimplement the forward pass from the safetensors header and the config — the architecture is simple enough to reconstruct — but that is reimplementation, not adoption, and it produces an unvalidated encoder with no reference outputs to check against.
Flash is missing verification. It runs, but no independent party has scored it, and the medical claims are vendor assertions. A hospital evaluating it would need its own labelled cases and its own reviewers, plus the multi-GPU hardware to serve the model while doing so.
Together they are missing the connective tissue: no shared embedding space, no documented projection from Pulse vectors into flash's token stream, no joint benchmark, and no reference pipeline. The family name suggests a stack. The repositories describe two components with a gap between them.
Was Sie heute ausführen können
Flash: yes, if you have four H20s or eight H200s and a serving stack that supports the Bailing hybrid architecture. The SGLang image and vLLM recipe are published, and the MXFP4 checkpoint lowers the hardware floor to 292 GB. What you cannot do is confirm its medical quality without building your own evaluation.
Pulse: not end to end. You can download 63.8 MiB of weights and read a genuinely informative header. You cannot call the model through any published interface. Until a package ships, the encoder is a specification rather than a tool.
OrcaRouter hosts neither model, so neither half of this pipeline is available through us today. The relevant product point is narrower: when a team does assemble a multi-stage clinical stack, the serving layer is where the pieces either compose or fail — one endpoint for hosted models, routing that can send a request to the right stage, and automatic failover when a stage degrades. That is plumbing for after the experiment, not a substitute for it.
The honest summary is that inclusionAI has published two interesting components of a medical AI system and no system. The text model is the mature half; the signal encoder is the promising half; the integration is neither. Anyone who needs a working clinical pipeline today should treat this pair as a design reference, not a deployment plan — and anyone who builds the missing seam will have done the part the vendor has not.

