v0.7.14

Native install (Apple Silicon / homelab)

uv tool install mship
mship bootstrap --metal
mship deploy --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/modelship-ai/modelship:0.7.14

CUDA (GPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.14-cuda

CPU (CPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.14-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/modelship-ai/charts/modelship \
  --version 0.7.14 \
  -f values.yaml

Running an image

The images take the same subcommands as the mship CLI:

docker run --rm --shm-size=8g \
  -v ./models.yaml:/modelship/config/models.yaml \
  -v ./models-cache:/.cache \
  -p 8000:8000 \
  ghcr.io/modelship-ai/modelship:0.7.14-cpu deploy

Configuration

Create a models.yaml file and mount it at /modelship/config/models.yaml. Example configs for various GPU sizes are included in the image under /modelship/config/examples/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• Fix/gemma4 vllm support by @alez007 in #180
• Fix/llama server sliding window preflight by @alez007 in #181
• Fix/mla preflight sizing by @alez007 in #183
• Fix/hybrid preflight sizing by @alez007 in #184
• fix: log silent ErrorResponse paths and validate tool_choice at the s… by @alez007 in #185
• Feat/devcontainer llama server provisioning by @alez007 in #186
• fix: emit an SSE error chunk on mid-stream failures instead of aborting by @alez007 in #187
• feat: aggregate HF model download progress and add a heartbeat by @alez007 in #188
• Feat/cli nested model flags by @alez007 in #189

Full Changelog: v0.7.13...v0.7.14

v0.7.13

Native install (Apple Silicon / homelab)

uv tool install mship
mship bootstrap --metal
mship deploy --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/modelship-ai/modelship:0.7.13

CUDA (GPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.13-cuda

CPU (CPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.13-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/modelship-ai/charts/modelship \
  --version 0.7.13 \
  -f values.yaml

Running an image

The images take the same subcommands as the mship CLI:

docker run --rm --shm-size=8g \
  -v ./models.yaml:/modelship/config/models.yaml \
  -v ./models-cache:/.cache \
  -p 8000:8000 \
  ghcr.io/modelship-ai/modelship:0.7.13-cpu deploy

Configuration

Create a models.yaml file and mount it at /modelship/config/models.yaml. Example configs for various GPU sizes are included in the image under /modelship/config/examples/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• feat: split provisioning into mship bootstrap by @alez007 in #174
• Feat/docker native parity by @alez007 in #175
• feat: gate bootstrap --cuda on the CUDA toolchain, not the GPU by @alez007 in #176
• Refactor/ray free config schema by @alez007 in #177
• Fix/gateway route prefix by @alez007 in #178
• feat: deploy a single model from CLI flags instead of models.yaml by @alez007 in #179

Full Changelog: v0.7.12...v0.7.13

v0.7.12

Native install (Apple Silicon / homelab)

uv tool install "mship[metal]"
mship deploy --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/modelship-ai/modelship:0.7.12

CUDA (GPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.12-cuda

CPU (CPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.12-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/modelship-ai/charts/modelship \
  --version 0.7.12 \
  -f values.yaml

Configuration

Create a models.yaml file and mount it at /modelship/config/models.yaml. Example configs for various GPU sizes are included in the image under /modelship/config/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• Feat/mship cpu native install by @alez007 in #168
• ci: add Linux CUDA llama.cpp backend build workflow by @alez007 in #169
• Ci/llama cpp cuda publish guard by @alez007 in #170
• Ci/llama cpp unified build by @alez007 in #171
• Feat/native cuda install by @alez007 in #172
• Feat/bootstrapper by @alez007 in #173

Full Changelog: v0.7.11...v0.7.12

v0.7.11

Native install (Apple Silicon / homelab)

uv tool install "mship[metal]"
mship deploy --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/modelship-ai/modelship:0.7.11

CUDA (GPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.11-cuda

CPU (CPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.11-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/modelship-ai/charts/modelship \
  --version 0.7.11 \
  -f values.yaml

Configuration

Create a models.yaml file and mount it at /modelship/config/models.yaml. Example configs for various GPU sizes are included in the image under /modelship/config/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• fix: drop CoreML support from sherpa_onnx loader, CPU only by @alez007 in #165
• Fix/throttle download progress logs by @alez007 in #166
• fix: stop whispercpp loader's native/download logs bleeding into console by @alez007 in #167

Full Changelog: v0.7.10...v0.7.11

v0.7.10

Native install (Apple Silicon / homelab)

uv tool install "mship[metal]"
mship deploy --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/modelship-ai/modelship:0.7.10

CUDA (GPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.10-cuda

CPU (CPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.10-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/modelship-ai/charts/modelship \
  --version 0.7.10 \
  -f values.yaml

Configuration

Create a models.yaml file and mount it at /modelship/config/models.yaml. Example configs for various GPU sizes are included in the image under /modelship/config/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• Feat/fractional gpu sharing by @alez007 in #164

Full Changelog: v0.7.9...v0.7.10

v0.7.9

Native install (Apple Silicon / homelab)

uv tool install "mship[metal]"
mship deploy --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/modelship-ai/modelship:0.7.9

CUDA (GPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.9-cuda

CPU (CPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.9-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/modelship-ai/charts/modelship \
  --version 0.7.9 \
  -f values.yaml

Configuration

Create a models.yaml file and mount it at /modelship/config/models.yaml. Example configs for various GPU sizes are included in the image under /modelship/config/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• fix: cut library log noise, prefix and de-duplicate startup log lines by @alez007 in #154
• fix: close vLLM cross-identity prefix-cache timing side channel by @alez007 in #155
• fix: refresh stale and broken model references in example configs by @alez007 in #157
• fix: crash the replica actor when a loader's backend dies unexpectedly by @alez007 in #158
• fix: raise onnxruntime-gpu floor to 1.28.0 for CUDA 13 compatibility by @alez007 in #159
• feat: promote whispercpp to a first-class in-process loader by @alez007 in #160
• refactor: remove the whispercpp plugin, superseded by loader: whispercpp by @alez007 in #161
• feat: add sherpa_onnx TTS loader (kokoro, CPU/CoreML, streaming) by @alez007 in #162
• refactor: remove the plugin system by @alez007 in #163

Full Changelog: v0.7.8...v0.7.9

v0.7.8

Native install (Apple Silicon / homelab)

uv tool install "mship[metal]"
mship deploy --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/modelship-ai/modelship:0.7.8

CUDA (GPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.8-cuda

CPU (CPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.8-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/modelship-ai/charts/modelship \
  --version 0.7.8 \
  -f values.yaml

Configuration

Create a models.yaml file and mount it at /modelship/config/models.yaml. Example configs for various GPU sizes are included in the image under /modelship/config/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• Refactor/remove round robin by @alez007 in #152
• feat: capability-aware Ray scheduling for loaders and accelerators by @alez007 in #153

Full Changelog: v0.7.7...v0.7.8

v0.7.7

Native install (Apple Silicon / homelab)

uv tool install "mship[metal]"
mship deploy --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/modelship-ai/modelship:0.7.7

CUDA (GPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.7-cuda

CPU (CPU node)

docker pull ghcr.io/modelship-ai/modelship:0.7.7-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/modelship-ai/charts/modelship \
  --version 0.7.7 \
  -f values.yaml

Configuration

Create a models.yaml file and mount it at /modelship/config/models.yaml. Example configs for various GPU sizes are included in the image under /modelship/config/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• docs: point repository references at the modelship-ai org by @alez007 in #151

Full Changelog: v0.7.6...v0.7.7

v0.7.6

Native install (Apple Silicon / homelab)

uv tool install "mship[metal]"
mship deploy --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/alez007/modelship:0.7.6

CUDA (GPU node)

docker pull ghcr.io/alez007/modelship:0.7.6-cuda

CPU (CPU node)

docker pull ghcr.io/alez007/modelship:0.7.6-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/alez007/charts/modelship \
  --version 0.7.6 \
  -f values.yaml

Configuration

Create a models.yaml file and mount it at /modelship/config/models.yaml. Example configs for various GPU sizes are included in the image under /modelship/config/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• chore: add CLA, CODEOWNERS, and governance docs for contributors by @alez007 in #146
• Fix/mcp policy shape validation by @alez007 in #147
• Chore/upgrade vllm 0.26.0 by @alez007 in #148
• fix: auto-detect free host RAM for Ray node memory sizing by @alez007 in #149
• chore: bump llama.cpp to b10200 by @github-actions[bot] in #150

Full Changelog: v0.7.5...v0.7.6

v0.7.5

Native install (Apple Silicon / homelab)

uv tool install "mship[metal]"
mship deploy --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/alez007/modelship:0.7.5

CUDA (GPU node)

docker pull ghcr.io/alez007/modelship:0.7.5-cuda

CPU (CPU node)

docker pull ghcr.io/alez007/modelship:0.7.5-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/alez007/charts/modelship \
  --version 0.7.5 \
  -f values.yaml

Configuration

Create a models.yaml file and mount it at /modelship/config/models.yaml. Example configs for various GPU sizes are included in the image under /modelship/config/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• feat: add background mode (Phase E1) to /v1/responses by @alez007 in #143
• feat: add background+stream live tailing and resume (Phase E2) to /v1… by @alez007 in #144
• feat: server-side MCP tool execution on /v1/responses by @alez007 in #145

Full Changelog: v0.7.4...v0.7.5