Skip to content
AI Atlas
ModelAnnouncedClosed / proprietary

CARE-X

Microsoft

Open in Graph
data quality56

Updated 2 h ago · first seen 11 Sept 2026

model_01M296QTDT4HJPWMMAAZ4M4VQD

Overview

Identity

Canonical model
Yesidentity confidence: highOne row per real model release. Artifacts (checkpoints, quantisations, conversions) and folded evaluation variants point here.
Official checkpoints
None recorded — closed weightsofficial_checkpoints = hf_repo identifiers carried by the model itself; artifacts are separate entities pointing here through canonical_id.
Artifacts
None recordedSeparate entities (checkpoint · quantization · conversion · packaging) pointing to this model through canonical_id.
Provider deployments
None recorded
API aliases
NoneIdentifiers under which providers and evaluators refer to this model.
Folded evaluation variants
0Effort / thinking variants (…-high, …-non-reasoning) are result configurations of this model, not separate models. Their old URLs redirect here.

Openness

Closed / proprietary— weights not available; 7 dimensions unknown.

Weights are not available; the model is reachable only through an API or a product.

  • Weights

    No

  • Inference code

    —

  • Training code

    —

  • Training data

    —

  • Dataset

    —

  • Commercial use

    —

  • Redistribution

    —

  • Derivatives

    —

dimensions marked null are unknown, not false

Key facts

Release date

Source:Microsoft Research — publications & blogT2observed 12 h agomediumLLM-extracted

Status

Source:Microsoft Research — publications & blogT2observed 12 h agomediumLLM-extracted

Architecture

Architecture

Source:Microsoft Research — publications & blogT2observed 12 h agomediumLLM-extracted

Parameters

Source:Microsoft Research — publications & blogT2observed 12 h agomediumLLM-extracted

Capabilities

Modalities

Modalities
imagetext
Input
imagetext
Output
text

Capabilities

  • Tool calling

    Yes

    Microsoft Research — publications & blog · T2

  • Structured output

    Yes

    Microsoft Research — publications & blog · T2

  • Reasoning

    Yes

    Microsoft Research — publications & blog · T2

  • Vision

    Yes

    Microsoft Research — publications & blog · T2

  • Audio

    No

    Microsoft Research — publications & blog · T2

  • Fine-tuning available

    Unavailable

No structured attributes yet.

Hardware fit37

Estimated

37 of 37 device × quantization combinations fit.

Run locally: your machine →
Estimated hardware fit
HardwareQuantizationDevice memoryEst. memoryFits
Apple M3 Ultra4bit—2.7 GB est.Yes
Mac Studio (Apple M5 Ultra)4bit—2.7 GB est.Yes
Apple M2 Ultra4bit—2.7 GB est.Yes
Apple M1 Ultra4bit—2.7 GB est.Yes
Apple M3 Max4bit—2.7 GB est.Yes
Apple M4 Max4bit—2.7 GB est.Yes
Mac Studio (Apple M5 Max)4bit—2.7 GB est.Yes
MacBook Pro (Apple M5 Max)4bit—2.7 GB est.Yes
Apple M2 Max4bit—2.7 GB est.Yes
Apple M1 Max4bit—2.7 GB est.Yes
Apple M4 Pro4bit—2.7 GB est.Yes
Mac mini (Apple M5 Pro)4bit—2.7 GB est.Yes
MacBook Pro (Apple M5 Pro)4bit—2.7 GB est.Yes
Apple M3 Pro4bit—2.7 GB est.Yes
Apple M1 Pro4bit—2.7 GB est.Yes
Apple M2 Pro4bit—2.7 GB est.Yes
Apple M44bit—2.7 GB est.Yes
iMac (Apple M4)4bit—2.7 GB est.Yes
Mac mini (Apple M6)4bit—2.7 GB est.Yes
MacBook Air (Apple M5)4bit—2.7 GB est.Yes
MacBook Pro (Apple M5)4bit—2.7 GB est.Yes
Apple M24bit—2.7 GB est.Yes
Apple M34bit—2.7 GB est.Yes
Apple M14bit—2.7 GB est.Yes
NVIDIA GeForce RTX 30904bit24 GB2.7 GB est.Yes
NVIDIA GeForce RTX 40904bit24 GB2.7 GB est.Yes
NVIDIA GeForce RTX 50904bit32 GB2.7 GB est.Yes
NVIDIA A100 80GB4bit80 GB2.7 GB est.Yes
NVIDIA H100 SXM4bit80 GB2.7 GB est.Yes
NVIDIA H100 NVL4bit94 GB2.7 GB est.Yes
NVIDIA DGX Spark4bit128 GB2.7 GB est.Yes
NVIDIA H2004bit141 GB2.7 GB est.Yes
NVIDIA H200 NVL4bit141 GB2.7 GB est.Yes
NVIDIA B2004bit180 GB2.7 GB est.Yes
AMD Instinct MI300X4bit192 GB2.7 GB est.Yes
AMD Instinct MI325X4bit256 GB2.7 GB est.Yes
NVIDIA DGX B2004bit1,440 GB2.7 GB est.Yes
Assumptions (7)
  • Estimated, not measured: weights = parameters × bytes/param × 1.15 runtime overhead (or the observed artifact file size when one is recorded).
  • bytes/param: 4bit = 0.5, 8bit = 1.0, fp16 = 2.0 (uniform quantization, no per-layer exceptions).
  • KV cache: 2 × layers × kv_heads × head_dim × 2 bytes × context × batch when the architecture is known; otherwise 0.5 GB per 8 192 tokens (× batch), independent of architecture (GQA/MLA models need less).
  • A model 'fits' when the estimate is at most the device memory minus 2 GB reserved for the OS and framework.
  • Mixture-of-experts models are estimated on total parameters (all experts must be resident); active parameters are ignored.
  • Device memory uses the largest configuration when several are listed (e.g. Apple silicon tiers).
  • Multi-GPU: device memories are summed; interconnect bandwidth, tensor-parallel replication and pipeline bubbles are not modelled.

Lineage

Open in Graph →
Explicit derived_from / fine_tuned_from / distilled_from relations stated by sources; artifacts collapsed by kind.
ANCESTORS 2Phi-4-mini-instruct (3.8B)Phi-4-mini-instruct (3.8B)Phi-4-mini-instruct (3.8B), S…Phi-4-mini-instruct (3.8B), SigLIP2-so400MCARE-X3.8B params · this modelCARE-X — 3.8B params · this model

Versions & Artifacts0

Version history

Opennessfirst observation only

11 Sept 2026current

Parametersfirst observation only

11 Sept 2026current

Statusfirst observation only

11 Sept 2026current

Each hop is a claim: click a value for its source, tier and observation time. Nothing is overwritten — a new observation closes the previous claim.

Artifacts 0

No artifact (checkpoint, quantisation, conversion or packaging) points to this model yet.

Change history17

Temporal, append-only claims: a new observation closes the previous claim instead of overwriting it. Rewind the record with the as-of picker.

Viewing AI Atlas as of 12 Aug 2026 — attributes exactly as the atlas knew them on that day; later corrections are not shown.

Back to today →

CARE-X was not yet in AI Atlas on 12 Aug 2026

First seen 11 Sept 2026. Nothing is inferred backwards: no attribute is shown for dates before the first observation.
17 claims · 16 properties

Release daterelease_date1

Claim history for Release date
ValueValid from → toStatusSourceConfidenceExtractor
11 Aug 2026 → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Statusstatus1

Claim history for Status
ValueValid from → toStatusSourceConfidenceExtractor
announced → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Opennessopenness1

Claim history for Openness
ValueValid from → toStatusSourceConfidenceExtractor
proprietary → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Architecturearchitecture1

Claim history for Architecture
ValueValid from → toStatusSourceConfidenceExtractor
SigLIP2-so400M vision encoder and Phi-4-mini-instruct (3.8B) language model connected through a lightweight adapter, with task-specific auxiliary heads for classification and visual grounding → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Parametersparameter_count1

Claim history for Parameters
ValueValid from → toStatusSourceConfidenceExtractor
3.8B → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Modalitiesmodalities1

Claim history for Modalities
ValueValid from → toStatusSourceConfidenceExtractor
image, text → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Input modalitiesmodalities_input1

Claim history for Input modalities
ValueValid from → toStatusSourceConfidenceExtractor
image, text → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Output modalitiesmodalities_output1

Claim history for Output modalities
ValueValid from → toStatusSourceConfidenceExtractor
text → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Audioaudio1

Claim history for Audio
ValueValid from → toStatusSourceConfidenceExtractor
No → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Reasoningreasoning1

Claim history for Reasoning
ValueValid from → toStatusSourceConfidenceExtractor
Yes → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Safety notessafety_notes1

Claim history for Safety notes
ValueValid from → toStatusSourceConfidenceExtractor
CARE-X is a research model and not a Microsoft product offering or medical device. It has not been cleared or approved by any regulatory authority and is not intended for clinical diagnosis, screening, or patient care. The results described below are retrospective research findings and do not establish the safety, effectiveness, or suitability of CARE-X for any clinical use. → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Structured outputstructured_output1

Claim history for Structured output
ValueValid from → toStatusSourceConfidenceExtractor
Yes → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Tool callingtool_calling1

Claim history for Tool calling
ValueValid from → toStatusSourceConfidenceExtractor
Yes → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Training data notestraining_data_notes2

Claim history for Training data notes
ValueValid from → toStatusSourceConfidenceExtractor
Trained on chest X-ray data including MIMIC-CXR, IU-Xray, CheXpert-Plus, ReXGradient, and Narayana Health clinical data. Validated on 1,047 de-identified chest radiographs from Narayana Health and 122 positive cases with CT-confirmed ground truth. → currentcurrentMicrosoft Research — publications & blogT2mediumllm
Trained on real-world Indian clinical data from Narayana Health, including rare ICU pathologies and CT-confirmed enlargement conditions. Used de-identified, retrospective clinical data under applicable institutional ethics review. → supersededMicrosoft Research — publications & blogT2mediumllm

Visionvision1

Claim history for Vision
ValueValid from → toStatusSourceConfidenceExtractor
Yes → currentcurrentMicrosoft Research — publications & blogT2mediumllm

Weights availableweights_available1

Claim history for Weights available
ValueValid from → toStatusSourceConfidenceExtractor
No → currentcurrentAI Atlas curated registry (YAML, versioned in git, every entry carries its source URL)T2highderived

Claims are temporal and append-only: a new observation closes the previous claim (valid_to) instead of overwriting it. Conflicting claims from different sources are kept side by side and flagged — never averaged. Methodology →

Provenance

Attributed facts

16

Source tiers

T216

Freshest observation

2 h ago

Conflicts

None

Source documents 1

Source documents
SourceDocumentTypeTierLast observedSnapshots
Microsoft Research — publications & blogmicrosoft.com/en-us/research/blog/introducing-care-x-towards-clinically-useful-radiology-vlms-with-auxiliary-supervision-reward-aligned-learning-and-tool-augmented-measurement newsT1· Official5 h ago2

Tier 1 = official/primary, 2 = quality secondary, 3 = community, 4 = unverified. Every snapshot is archived; see all sources and the methodology.

Data quality (56/100) measures how well AI Atlas knows this entity — completeness, primary-source ratio, freshness, conflicts — never how good the model is. Methodology →