v1.14.0-ce — Native Vision¶
Released: 2026-07-29 · Tag: v1.14.0-ce · Docker: synapcores/community:v1.14.0-ce
Your database can see now. v1.14.0-ce ships native, in-process multimodal vision — send an image, ask a question, get an answer, generated on your own hardware with no cloud key and no second service to run.
Native vision — POST /v1/multimodal/describe¶
A new REST endpoint interprets an image with a vision LLM, entirely inside the gateway process:
curl -X POST http://127.0.0.1:8080/v1/multimodal/describe \
-H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
-d '{
"prompt": "Describe this image in one sentence.",
"image": "<base64 image>",
"provider": "local",
"model": "llava"
}'
# → { "text": "…", "model": "llava" }
- Native + in-process. The vision model runs in the same process as your database via a safe
libmtmd/ LLaVA binding. No cloud round-trip, no external daemon, no API key. The image never leaves the box. - CPU-friendly. Validated on commodity, non-AVX-512 hardware — no GPU required.
- Bring your own model, or ours. Defaults to LLaVA; set
modelper request orAIDB_VISION_MODELto point at the vision model you prefer. Cloud providers (openai,anthropic,ollama) still work when configured.
Pull a vision model into the in-process registry once, then it's available with no daemon:
Why it matters¶
Vision used to mean shipping your images to someone else's server, per-call billing, and hoping
the data was allowed to leave. Now SynapCores does it locally, privately, on hardware you already
run — alongside the SQL, vectors, knowledge graph, native LLMs, and agents already in the engine.
Give an in-database AGENT_RUN workflow eyes; read scanned documents where they land; triage
inspection photos next to the rows that describe the asset.
Validation¶
On the published artifact: CI build of 4 Linux targets with the native libmtmd compile;
feature_validator 88/88; recipe certification 158/162, zero regressions vs v1.13.0.1-ce;
non-AVX-512 canary (i5-10400F) boots and runs all-minilm + llava with no illegal-instruction
(ZMM instructions cpuid-gated); native vision end-to-end described a test image in-process on
CPU on both the raw binary and the published Docker image.
Install¶
docker pull synapcores/community:v1.14.0-ce
# or:
curl -fsSL https://get.synapcores.com | SYNAPCORES_VERSION=v1.14.0-ce sh
Drop-in upgrade — no schema migration. The vision route is additive; the existing REST / WebSocket / SDK / MySQL-wire paths are unchanged.