diff --git a/spinifex/gateway/bedrock/handler.go b/spinifex/gateway/bedrock/handler.go
index fe827b4e3..51787261d 100644
--- a/spinifex/gateway/bedrock/handler.go
+++ b/spinifex/gateway/bedrock/handler.go
@@ -56,6 +56,17 @@ func WriteJSONResponse(w http.ResponseWriter, obj any) {
 	}
 }
 
+// WriteRawResponse writes body verbatim with contentType and a 200 status,
+// bypassing JSON marshaling. Used by InvokeModel, whose response is the
+// provider-native body rather than a struct.
+func WriteRawResponse(w http.ResponseWriter, body []byte, contentType string) {
+	w.Header().Set("Content-Type", contentType)
+	w.WriteHeader(http.StatusOK)
+	if _, err := w.Write(body); err != nil {
+		slog.Error("Failed to write bedrock raw response", "err", err)
+	}
+}
+
 // WriteJSONError emits the AWS REST-JSON error envelope with the given code,
 // message, and HTTP status.
 func WriteJSONError(w http.ResponseWriter, code, message string, httpStatus int) {
diff --git a/spinifex/gateway/bedrock/invoke.go b/spinifex/gateway/bedrock/invoke.go
new file mode 100644
index 000000000..d1872cdbb
--- /dev/null
+++ b/spinifex/gateway/bedrock/invoke.go
@@ -0,0 +1,80 @@
+package gateway_bedrock
+
+import (
+	"context"
+	"errors"
+	"strings"
+
+	"github.com/mulgadc/spinifex/spinifex/awserrors"
+)
+
+// InvokeAdapter translates a Bedrock InvokeModel raw request body into a
+// backend's native wire format and back, returning the response bytes
+// verbatim with their content-type. Unlike Provider (Converse), the wire
+// shape is per-family and is not unified by the gateway.
+type InvokeAdapter interface {
+	InvokeModel(ctx context.Context, modelID string, body []byte) (respBody []byte, contentType string, err error)
+}
+
+// InvokeRouter resolves a modelId to its catalog entry and dispatches to the
+// matching InvokeAdapter, resolving self-host endpoints and provider
+// credentials as needed.
+type InvokeRouter struct {
+	resolver         CredentialResolver
+	endpointResolver EndpointResolver
+}
+
+// NewInvokeRouter constructs an InvokeRouter. A nil resolver or
+// endpointResolver falls back to a resolver/resolver that finds nothing, so
+// an InvokeRouter is always safe to use even before the real stores are
+// wired in.
+func NewInvokeRouter(resolver CredentialResolver, endpointResolver EndpointResolver) *InvokeRouter {
+	if resolver == nil {
+		resolver = NoopCredentialResolver
+	}
+	if endpointResolver == nil {
+		endpointResolver = NewStaticEndpointResolver(nil)
+	}
+	return &InvokeRouter{resolver: resolver, endpointResolver: endpointResolver}
+}
+
+// InvokeModel routes modelID to its family adapter via the catalog. Unknown
+// modelIds and unresolvable vendors return ResourceNotFoundException; a
+// vendor with no resolvable credential returns AccessDeniedException.
+func (rt *InvokeRouter) InvokeModel(ctx context.Context, accountID, modelID string, body []byte) ([]byte, string, error) {
+	entry, ok := lookupCatalogEntry(modelID)
+	if !ok {
+		return nil, "", errors.New(awserrors.ErrorResourceNotFoundException)
+	}
+
+	var a InvokeAdapter
+	switch {
+	case entry.Provider == tierSelfHost:
+		a = newLlamaInvokeAdapter(rt.endpointResolver)
+	case strings.HasPrefix(entry.Provider, providerPrefix):
+		switch strings.TrimPrefix(entry.Provider, providerPrefix) {
+		case vendorAnthropic:
+			key, ok, err := rt.resolver.Resolve(ctx, accountID, vendorAnthropic)
+			if err != nil {
+				return nil, "", err
+			}
+			if !ok {
+				return nil, "", errors.New(awserrors.ErrorAccessDeniedException)
+			}
+			a = newAnthropicInvokeAdapter(key)
+		default:
+			return nil, "", errors.New(awserrors.ErrorResourceNotFoundException)
+		}
+	default:
+		return nil, "", errors.New(awserrors.ErrorResourceNotFoundException)
+	}
+
+	return a.InvokeModel(ctx, modelID, body)
+}
+
+// InvokeModel is the bedrock-runtime InvokeModel entry point used by the
+// gateway route table. resolver and endpointResolver may be nil;
+// NewInvokeRouter supplies no-op fallbacks.
+func InvokeModel(ctx context.Context, accountID, modelID string, body []byte, resolver CredentialResolver, endpointResolver EndpointResolver) ([]byte, string, error) {
+	return NewInvokeRouter(resolver, endpointResolver).InvokeModel(ctx, accountID, modelID, body)
+}
diff --git a/spinifex/gateway/bedrock/invoke_anthropic.go b/spinifex/gateway/bedrock/invoke_anthropic.go
new file mode 100644
index 000000000..62ebcc6e6
--- /dev/null
+++ b/spinifex/gateway/bedrock/invoke_anthropic.go
@@ -0,0 +1,102 @@
+package gateway_bedrock
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"log/slog"
+	"net/http"
+
+	"github.com/mulgadc/spinifex/spinifex/awserrors"
+)
+
+// anthropicInvokeAdapter forwards Bedrock-native Claude InvokeModel bodies to
+// the Anthropic Messages API almost as-is. httpClient and baseURL are
+// injectable for tests.
+type anthropicInvokeAdapter struct {
+	httpClient *http.Client
+	baseURL    string
+}
+
+func newAnthropicInvokeAdapterClient() *anthropicInvokeAdapter {
+	return &anthropicInvokeAdapter{
+		httpClient: &http.Client{Timeout: providerHTTPTimeout},
+		baseURL:    anthropicDefaultBaseURL,
+	}
+}
+
+// boundAnthropicInvokeAdapter adapts anthropicInvokeAdapter to InvokeAdapter
+// by baking in a resolved per-account (or platform-default) API key.
+type boundAnthropicInvokeAdapter struct {
+	inner  *anthropicInvokeAdapter
+	apiKey string
+}
+
+var _ InvokeAdapter = (*boundAnthropicInvokeAdapter)(nil)
+
+func (b *boundAnthropicInvokeAdapter) InvokeModel(ctx context.Context, modelID string, body []byte) ([]byte, string, error) {
+	return b.inner.InvokeModel(ctx, modelID, body, b.apiKey)
+}
+
+// newAnthropicInvokeAdapter returns an InvokeAdapter that forwards to the
+// Anthropic Messages API with apiKey.
+func newAnthropicInvokeAdapter(apiKey string) InvokeAdapter {
+	return &boundAnthropicInvokeAdapter{inner: newAnthropicInvokeAdapterClient(), apiKey: apiKey}
+}
+
+// InvokeModel rewrites the Bedrock Claude InvokeModel body (drops the
+// Bedrock-only anthropic_version field, injects model since Anthropic's API
+// carries it in the body rather than the URL) and posts it to the Anthropic
+// Messages API, returning the response verbatim.
+func (a *anthropicInvokeAdapter) InvokeModel(ctx context.Context, modelID string, body []byte, apiKey string) ([]byte, string, error) {
+	var fields map[string]json.RawMessage
+	if err := json.Unmarshal(body, &fields); err != nil {
+		slog.Error("anthropic invoke: failed to parse request body", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorValidationException)
+	}
+	delete(fields, "anthropic_version")
+
+	modelJSON, err := json.Marshal(anthropicModelID(modelID))
+	if err != nil {
+		slog.Error("anthropic invoke: failed to marshal model id", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorInternalError)
+	}
+	fields["model"] = modelJSON
+
+	reqBody, err := json.Marshal(fields)
+	if err != nil {
+		slog.Error("anthropic invoke: failed to marshal request", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorInternalError)
+	}
+
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, a.baseURL+anthropicMessagesPath, bytes.NewReader(reqBody))
+	if err != nil {
+		slog.Error("anthropic invoke: failed to build request", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorInternalError)
+	}
+	httpReq.Header.Set("X-Api-Key", apiKey)
+	httpReq.Header.Set("Anthropic-Version", anthropicAPIVersion)
+	httpReq.Header.Set("Content-Type", "application/json")
+
+	resp, err := a.httpClient.Do(httpReq)
+	if err != nil {
+		slog.Error("anthropic invoke: request failed", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorServiceUnavailableException)
+	}
+	defer resp.Body.Close()
+
+	respBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		slog.Error("anthropic invoke: failed to read response body", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorServiceUnavailableException)
+	}
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		slog.Error("anthropic invoke: upstream error", "model", modelID, "status", resp.StatusCode, "body", string(respBody))
+		return nil, "", errors.New(mapUpstreamStatus(resp.StatusCode))
+	}
+
+	return respBody, "application/json", nil
+}
diff --git a/spinifex/gateway/bedrock/invoke_anthropic_test.go b/spinifex/gateway/bedrock/invoke_anthropic_test.go
new file mode 100644
index 000000000..6c9e452a9
--- /dev/null
+++ b/spinifex/gateway/bedrock/invoke_anthropic_test.go
@@ -0,0 +1,97 @@
+package gateway_bedrock
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/mulgadc/spinifex/spinifex/awserrors"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestAnthropicInvokeAdapter_InvokeModel_RewritesRequestAndReturnsBodyVerbatim(t *testing.T) {
+	canned := `{"id":"msg_123","content":[{"type":"text","text":"Hello there"}],"stop_reason":"end_turn"}`
+
+	var captured map[string]json.RawMessage
+	var capturedAPIKey, capturedAnthropicVersion, capturedContentType string
+
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		capturedAPIKey = r.Header.Get("X-Api-Key")
+		capturedAnthropicVersion = r.Header.Get("Anthropic-Version")
+		capturedContentType = r.Header.Get("Content-Type")
+		require.NoError(t, json.NewDecoder(r.Body).Decode(&captured))
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(canned))
+	}))
+	defer ts.Close()
+
+	a := &anthropicInvokeAdapter{httpClient: ts.Client(), baseURL: ts.URL}
+
+	reqBody := []byte(`{"anthropic_version":"bedrock-2023-05-31","max_tokens":256,"messages":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`)
+
+	respBody, contentType, err := a.InvokeModel(context.Background(), "anthropic.claude-3-5-sonnet-20240620-v1:0", reqBody, "sk-test-key")
+	require.NoError(t, err)
+
+	assert.Equal(t, "sk-test-key", capturedAPIKey)
+	assert.Equal(t, anthropicAPIVersion, capturedAnthropicVersion)
+	assert.Equal(t, "application/json", capturedContentType)
+
+	_, hasVersion := captured["anthropic_version"]
+	assert.False(t, hasVersion, "anthropic_version must be dropped from the forwarded request")
+
+	var model string
+	require.NoError(t, json.Unmarshal(captured["model"], &model))
+	assert.Equal(t, "claude-3-5-sonnet-20240620", model)
+
+	require.Contains(t, captured, "max_tokens")
+	require.Contains(t, captured, "messages")
+
+	assert.Equal(t, "application/json", contentType)
+	assert.JSONEq(t, canned, string(respBody))
+}
+
+func TestAnthropicInvokeAdapter_InvokeModel_401ReturnsAccessDenied(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusUnauthorized)
+		_, _ = w.Write([]byte(`{"error": {"message": "invalid x-api-key"}}`))
+	}))
+	defer ts.Close()
+
+	a := &anthropicInvokeAdapter{httpClient: ts.Client(), baseURL: ts.URL}
+
+	_, _, err := a.InvokeModel(context.Background(), "anthropic.claude-3-5-sonnet-20240620-v1:0", []byte(`{"anthropic_version":"bedrock-2023-05-31","messages":[]}`), "bad-key")
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorAccessDeniedException, err.Error())
+}
+
+// TestBoundAnthropicInvokeAdapter_InvokeModel exercises newAnthropicInvokeAdapter
+// and boundAnthropicInvokeAdapter.InvokeModel (the InvokeAdapter used by
+// InvokeRouter for the "provider:anthropic" tier). newAnthropicInvokeAdapter
+// hardcodes the real Anthropic base URL, so — same package as production code
+// — the test reaches into the unexported inner client to redirect it at an
+// httptest stub instead of the network.
+func TestBoundAnthropicInvokeAdapter_InvokeModel(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		assert.Equal(t, "sk-bound-test", r.Header.Get("X-Api-Key"))
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{"content":[{"type":"text","text":"bound response"}],"stop_reason":"end_turn"}`))
+	}))
+	defer ts.Close()
+
+	adapter := newAnthropicInvokeAdapter("sk-bound-test")
+	bound, ok := adapter.(*boundAnthropicInvokeAdapter)
+	require.True(t, ok)
+	require.NotNil(t, bound.inner)
+	bound.inner.httpClient = ts.Client()
+	bound.inner.baseURL = ts.URL
+
+	respBody, contentType, err := adapter.InvokeModel(context.Background(), "anthropic.claude-3-5-sonnet-20240620-v1:0", []byte(`{"anthropic_version":"bedrock-2023-05-31","messages":[]}`))
+	require.NoError(t, err)
+	assert.Equal(t, "application/json", contentType)
+	assert.Contains(t, string(respBody), "bound response")
+}
diff --git a/spinifex/gateway/bedrock/invoke_llama.go b/spinifex/gateway/bedrock/invoke_llama.go
new file mode 100644
index 000000000..ba4313151
--- /dev/null
+++ b/spinifex/gateway/bedrock/invoke_llama.go
@@ -0,0 +1,166 @@
+package gateway_bedrock
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"log/slog"
+	"net/http"
+
+	"github.com/mulgadc/spinifex/spinifex/awserrors"
+)
+
+const llamaCompletionsPath = "/v1/completions"
+
+// llamaInvokeRequest is the Bedrock-native Llama InvokeModel request body.
+type llamaInvokeRequest struct {
+	Prompt      string   `json:"prompt"`
+	MaxGenLen   int      `json:"max_gen_len,omitempty"`
+	Temperature *float64 `json:"temperature,omitempty"`
+	TopP        *float64 `json:"top_p,omitempty"`
+}
+
+// llamaInvokeResponse is the Bedrock-native Llama InvokeModel response body.
+type llamaInvokeResponse struct {
+	Generation           string `json:"generation"`
+	PromptTokenCount     int    `json:"prompt_token_count"`
+	GenerationTokenCount int    `json:"generation_token_count"`
+	StopReason           string `json:"stop_reason"`
+}
+
+// llamaCompletionsRequest is the OpenAI /v1/completions request vLLM serves.
+type llamaCompletionsRequest struct {
+	Model       string   `json:"model"`
+	Prompt      string   `json:"prompt"`
+	MaxTokens   int      `json:"max_tokens,omitempty"`
+	Temperature *float64 `json:"temperature,omitempty"`
+	TopP        *float64 `json:"top_p,omitempty"`
+}
+
+type llamaCompletionsChoice struct {
+	Text         string `json:"text"`
+	FinishReason string `json:"finish_reason"`
+}
+
+type llamaCompletionsUsage struct {
+	PromptTokens     int `json:"prompt_tokens"`
+	CompletionTokens int `json:"completion_tokens"`
+}
+
+type llamaCompletionsResponse struct {
+	Choices []llamaCompletionsChoice `json:"choices"`
+	Usage   llamaCompletionsUsage    `json:"usage"`
+}
+
+// llamaFinishReasons maps OpenAI finish_reason values to Bedrock Llama's
+// stop_reason values. Unrecognised values fall back to "stop".
+var llamaFinishReasons = map[string]string{
+	"stop":   "stop",
+	"length": "length",
+}
+
+func mapLlamaStopReason(reason string) string {
+	if mapped, ok := llamaFinishReasons[reason]; ok {
+		return mapped
+	}
+	return "stop"
+}
+
+// llamaInvokeAdapter serves Bedrock-native Llama InvokeModel requests over a
+// self-hosted vLLM OpenAI-completions endpoint. httpClient and
+// endpointResolver are injectable for tests.
+type llamaInvokeAdapter struct {
+	endpointResolver EndpointResolver
+	httpClient       *http.Client
+}
+
+var _ InvokeAdapter = (*llamaInvokeAdapter)(nil)
+
+func newLlamaInvokeAdapter(endpointResolver EndpointResolver) *llamaInvokeAdapter {
+	return &llamaInvokeAdapter{
+		endpointResolver: endpointResolver,
+		httpClient:       &http.Client{Timeout: providerHTTPTimeout},
+	}
+}
+
+// InvokeModel resolves modelID's endpoint, translates the Bedrock-native
+// Llama request to an OpenAI completions request, calls the endpoint, and
+// translates the response back to the Bedrock Llama shape.
+func (a *llamaInvokeAdapter) InvokeModel(ctx context.Context, modelID string, body []byte) ([]byte, string, error) {
+	baseURL, ok, err := a.endpointResolver.Endpoint(ctx, modelID)
+	if err != nil {
+		slog.Error("llama invoke: endpoint resolution failed", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorServiceUnavailableException)
+	}
+	if !ok {
+		return nil, "", errors.New(awserrors.ErrorModelNotReadyException)
+	}
+
+	var lr llamaInvokeRequest
+	if err := json.Unmarshal(body, &lr); err != nil {
+		slog.Error("llama invoke: failed to parse request body", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorValidationException)
+	}
+
+	reqBody, err := json.Marshal(llamaCompletionsRequest{
+		Model:       modelID,
+		Prompt:      lr.Prompt,
+		MaxTokens:   lr.MaxGenLen,
+		Temperature: lr.Temperature,
+		TopP:        lr.TopP,
+	})
+	if err != nil {
+		slog.Error("llama invoke: failed to marshal request", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorInternalError)
+	}
+
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, baseURL+llamaCompletionsPath, bytes.NewReader(reqBody))
+	if err != nil {
+		slog.Error("llama invoke: failed to build request", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorInternalError)
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+
+	resp, err := a.httpClient.Do(httpReq)
+	if err != nil {
+		slog.Error("llama invoke: request failed", "model", modelID, "endpoint", baseURL, "err", err)
+		return nil, "", errors.New(awserrors.ErrorServiceUnavailableException)
+	}
+	defer resp.Body.Close()
+
+	respBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		slog.Error("llama invoke: failed to read response body", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorServiceUnavailableException)
+	}
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		slog.Error("llama invoke: upstream error", "model", modelID, "status", resp.StatusCode, "body", string(respBody))
+		return nil, "", errors.New(mapUpstreamStatus(resp.StatusCode))
+	}
+
+	var cr llamaCompletionsResponse
+	if err := json.Unmarshal(respBody, &cr); err != nil {
+		slog.Error("llama invoke: failed to parse response", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorModelErrorException)
+	}
+	if len(cr.Choices) == 0 {
+		return nil, "", errors.New(awserrors.ErrorModelErrorException)
+	}
+	choice := cr.Choices[0]
+
+	out, err := json.Marshal(llamaInvokeResponse{
+		Generation:           choice.Text,
+		PromptTokenCount:     cr.Usage.PromptTokens,
+		GenerationTokenCount: cr.Usage.CompletionTokens,
+		StopReason:           mapLlamaStopReason(choice.FinishReason),
+	})
+	if err != nil {
+		slog.Error("llama invoke: failed to marshal response", "model", modelID, "err", err)
+		return nil, "", errors.New(awserrors.ErrorInternalError)
+	}
+
+	return out, "application/json", nil
+}
diff --git a/spinifex/gateway/bedrock/invoke_llama_test.go b/spinifex/gateway/bedrock/invoke_llama_test.go
new file mode 100644
index 000000000..621beb560
--- /dev/null
+++ b/spinifex/gateway/bedrock/invoke_llama_test.go
@@ -0,0 +1,101 @@
+package gateway_bedrock
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/mulgadc/spinifex/spinifex/awserrors"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestLlamaInvokeAdapter_InvokeModel_MapsRequestAndResponse(t *testing.T) {
+	var captured llamaCompletionsRequest
+
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		require.NoError(t, json.NewDecoder(r.Body).Decode(&captured))
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{
+			"choices": [{"text": "hi there", "finish_reason": "stop"}],
+			"usage": {"prompt_tokens": 8, "completion_tokens": 3}
+		}`))
+	}))
+	defer ts.Close()
+
+	modelID := "meta.llama3-70b-instruct-v1:0"
+	a := newLlamaInvokeAdapter(NewStaticEndpointResolver(map[string]string{modelID: ts.URL}))
+	a.httpClient = ts.Client()
+
+	temp := 0.5
+	reqBody, err := json.Marshal(llamaInvokeRequest{Prompt: "hello", MaxGenLen: 128, Temperature: &temp})
+	require.NoError(t, err)
+
+	respBody, contentType, err := a.InvokeModel(context.Background(), modelID, reqBody)
+	require.NoError(t, err)
+	assert.Equal(t, "application/json", contentType)
+
+	assert.Equal(t, modelID, captured.Model)
+	assert.Equal(t, "hello", captured.Prompt)
+	assert.Equal(t, 128, captured.MaxTokens)
+	require.NotNil(t, captured.Temperature)
+	assert.Equal(t, 0.5, *captured.Temperature)
+
+	var out llamaInvokeResponse
+	require.NoError(t, json.Unmarshal(respBody, &out))
+	assert.Equal(t, "hi there", out.Generation)
+	assert.Equal(t, 8, out.PromptTokenCount)
+	assert.Equal(t, 3, out.GenerationTokenCount)
+	assert.Equal(t, "stop", out.StopReason)
+}
+
+func TestLlamaInvokeAdapter_InvokeModel_MapsLengthFinishReason(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{
+			"choices": [{"text": "truncated", "finish_reason": "length"}],
+			"usage": {"prompt_tokens": 1, "completion_tokens": 1}
+		}`))
+	}))
+	defer ts.Close()
+
+	modelID := "meta.llama3-70b-instruct-v1:0"
+	a := newLlamaInvokeAdapter(NewStaticEndpointResolver(map[string]string{modelID: ts.URL}))
+	a.httpClient = ts.Client()
+
+	respBody, _, err := a.InvokeModel(context.Background(), modelID, []byte(`{"prompt":"hello"}`))
+	require.NoError(t, err)
+
+	var out llamaInvokeResponse
+	require.NoError(t, json.Unmarshal(respBody, &out))
+	assert.Equal(t, "length", out.StopReason)
+}
+
+func TestLlamaInvokeAdapter_InvokeModel_EmptyChoicesReturnsModelError(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{"choices": [], "usage": {"prompt_tokens": 1, "completion_tokens": 0}}`))
+	}))
+	defer ts.Close()
+
+	modelID := "meta.llama3-70b-instruct-v1:0"
+	a := newLlamaInvokeAdapter(NewStaticEndpointResolver(map[string]string{modelID: ts.URL}))
+	a.httpClient = ts.Client()
+
+	_, _, err := a.InvokeModel(context.Background(), modelID, []byte(`{"prompt":"hello"}`))
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorModelErrorException, err.Error())
+}
+
+func TestLlamaInvokeAdapter_InvokeModel_UnresolvedEndpointReturnsModelNotReady(t *testing.T) {
+	a := newLlamaInvokeAdapter(NewStaticEndpointResolver(nil))
+
+	_, _, err := a.InvokeModel(context.Background(), "meta.llama3-70b-instruct-v1:0", []byte(`{"prompt":"hello"}`))
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorModelNotReadyException, err.Error())
+}
diff --git a/spinifex/gateway/bedrock/invoke_test.go b/spinifex/gateway/bedrock/invoke_test.go
new file mode 100644
index 000000000..31b4b5142
--- /dev/null
+++ b/spinifex/gateway/bedrock/invoke_test.go
@@ -0,0 +1,84 @@
+package gateway_bedrock
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/mulgadc/spinifex/spinifex/awserrors"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestInvokeRouter_SelfHostSuccess(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{
+			"choices": [{"text": "hi", "finish_reason": "stop"}],
+			"usage": {"prompt_tokens": 1, "completion_tokens": 1}
+		}`))
+	}))
+	defer ts.Close()
+
+	modelID := "meta.llama3-70b-instruct-v1:0"
+	rt := NewInvokeRouter(nil, NewStaticEndpointResolver(map[string]string{modelID: ts.URL}))
+
+	respBody, contentType, err := rt.InvokeModel(context.Background(), "000000000001", modelID, []byte(`{"prompt":"hello"}`))
+	require.NoError(t, err)
+	assert.Equal(t, "application/json", contentType)
+
+	var out llamaInvokeResponse
+	require.NoError(t, json.Unmarshal(respBody, &out))
+	assert.Equal(t, "hi", out.Generation)
+}
+
+func TestInvokeRouter_UnknownModelReturnsResourceNotFound(t *testing.T) {
+	rt := NewInvokeRouter(nil, nil)
+	_, _, err := rt.InvokeModel(context.Background(), "000000000001", "does.not-exist-v1:0", []byte(`{}`))
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorResourceNotFoundException, err.Error())
+}
+
+func TestInvokeRouter_AnthropicNoCredentialReturnsAccessDenied(t *testing.T) {
+	rt := NewInvokeRouter(stubCredentialResolver{ok: false}, nil)
+	_, _, err := rt.InvokeModel(context.Background(), "000000000001", "anthropic.claude-3-5-sonnet-20240620-v1:0", []byte(`{}`))
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorAccessDeniedException, err.Error())
+}
+
+func TestInvokeRouter_SelfHostNoEndpointReturnsModelNotReady(t *testing.T) {
+	rt := NewInvokeRouter(nil, nil)
+	_, _, err := rt.InvokeModel(context.Background(), "000000000001", "meta.llama3-70b-instruct-v1:0", []byte(`{"prompt":"hello"}`))
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorModelNotReadyException, err.Error())
+}
+
+func TestInvokeModel_PackageEntryPoint_SelfHostSuccess(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{
+			"choices": [{"text": "via package InvokeModel", "finish_reason": "stop"}],
+			"usage": {"prompt_tokens": 1, "completion_tokens": 1}
+		}`))
+	}))
+	defer ts.Close()
+
+	modelID := "meta.llama3-70b-instruct-v1:0"
+	respBody, contentType, err := InvokeModel(context.Background(), "000000000001", modelID, []byte(`{"prompt":"hello"}`), nil, NewStaticEndpointResolver(map[string]string{modelID: ts.URL}))
+	require.NoError(t, err)
+	assert.Equal(t, "application/json", contentType)
+
+	var out llamaInvokeResponse
+	require.NoError(t, json.Unmarshal(respBody, &out))
+	assert.Equal(t, "via package InvokeModel", out.Generation)
+}
+
+func TestInvokeModel_PackageEntryPoint_UnknownModel(t *testing.T) {
+	_, _, err := InvokeModel(context.Background(), "000000000001", "does.not-exist-v1:0", []byte(`{}`), nil, nil)
+	require.Error(t, err)
+	assert.Equal(t, awserrors.ErrorResourceNotFoundException, err.Error())
+}
diff --git a/spinifex/gateway/bedrock_request_test.go b/spinifex/gateway/bedrock_request_test.go
index ddcf49417..af445691c 100644
--- a/spinifex/gateway/bedrock_request_test.go
+++ b/spinifex/gateway/bedrock_request_test.go
@@ -35,6 +35,23 @@ func newVLLMStub(t *testing.T) *httptest.Server {
 	return ts
 }
 
+// newLlamaCompletionsStub stands up an httptest server that answers the
+// OpenAI completions wire, mirroring gateway/bedrock/invoke_llama_test.go's
+// stub, for the InvokeModel self-host path.
+func newLlamaCompletionsStub(t *testing.T) *httptest.Server {
+	t.Helper()
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{
+			"choices": [{"text": "hi from llama invoke", "finish_reason": "stop"}],
+			"usage": {"prompt_tokens": 5, "completion_tokens": 3}
+		}`))
+	}))
+	t.Cleanup(ts.Close)
+	return ts
+}
+
 // newBedrockRequestGateway builds a GatewayConfig with a real NATS connection
 // (satisfying Bedrock_Request/BedrockRuntime_Request's nil-check only — no
 // NATS subject handling is required for these routes), no IAMService (so
@@ -154,6 +171,25 @@ func TestBedrockRuntimeRequest_MalformedBodyReturnsValidationException(t *testin
 	assert.Equal(t, awserrors.ErrorValidationException, err.Error())
 }
 
+func TestBedrockRuntimeRequest_InvokeModel(t *testing.T) {
+	ts := newLlamaCompletionsStub(t)
+	gw := newBedrockRequestGateway(t, ts.URL)
+
+	body := `{"prompt":"hello","max_gen_len":128}`
+	req := bedrockRequestWithAccount(http.MethodPost, "/model/"+bedrockTestLlamaModelID+"/invoke", body)
+	w := httptest.NewRecorder()
+	require.NoError(t, gw.BedrockRuntime_Request(w, req))
+	require.Equal(t, http.StatusOK, w.Code)
+	assert.Equal(t, "application/json", w.Header().Get("Content-Type"))
+
+	assert.JSONEq(t, `{
+		"generation": "hi from llama invoke",
+		"prompt_token_count": 5,
+		"generation_token_count": 3,
+		"stop_reason": "stop"
+	}`, w.Body.String())
+}
+
 func TestBedrockRuntimeRequest_MissingAccountIDReturnsServerInternal(t *testing.T) {
 	ts := newVLLMStub(t)
 	gw := newBedrockRequestGateway(t, ts.URL)
diff --git a/spinifex/gateway/bedrockruntime.go b/spinifex/gateway/bedrockruntime.go
index e9d0c1bf0..95143fb55 100644
--- a/spinifex/gateway/bedrockruntime.go
+++ b/spinifex/gateway/bedrockruntime.go
@@ -29,7 +29,9 @@ type bedrockRuntimeRoute struct {
 // gw.bedrockEndpointResolver() over the configured pinned self-host endpoints.
 type bedrockRuntimeRouteHandler func(ctx context.Context, accountID string, params []string, body []byte, resolver gateway_bedrock.CredentialResolver, endpoints gateway_bedrock.EndpointResolver) (any, error)
 
-// bedrockRuntimeRoutes is the dispatch table.
+// bedrockRuntimeRoutes is the dispatch table. InvokeModel has no handler
+// function here: BedrockRuntime_Request special-cases its action to bypass
+// the JSON-marshaling dispatch below, since its response is raw bytes.
 var bedrockRuntimeRoutes = []bedrockRuntimeRoute{
 	{"POST", regexp.MustCompile(`^/model/([^/]+)/converse$`), "Converse",
 		func(ctx context.Context, acct string, p []string, b []byte, resolver gateway_bedrock.CredentialResolver, endpoints gateway_bedrock.EndpointResolver) (any, error) {
@@ -41,6 +43,7 @@ var bedrockRuntimeRoutes = []bedrockRuntimeRoute{
 			}
 			return gateway_bedrock.Converse(ctx, acct, p[0], input, resolver, endpoints)
 		}},
+	{"POST", regexp.MustCompile(`^/model/([^/]+)/invoke$`), "InvokeModel", nil},
 }
 
 // lookupBedrockRuntimeAction matches method+path against bedrockRuntimeRoutes,
@@ -103,6 +106,17 @@ func (gw *GatewayConfig) BedrockRuntime_Request(w http.ResponseWriter, r *http.R
 		return errors.New(awserrors.ErrorInvalidParameterValue)
 	}
 
+	// InvokeModel returns provider-native bytes, not a struct WriteJSONResponse
+	// could marshal, so it writes its own response body directly.
+	if action == "InvokeModel" {
+		respBody, contentType, err := gateway_bedrock.InvokeModel(r.Context(), accountID, params[0], body, gw.bedrockResolver(), gw.bedrockEndpointResolver())
+		if err != nil {
+			return err
+		}
+		gateway_bedrock.WriteRawResponse(w, respBody, contentType)
+		return nil
+	}
+
 	output, err := handler(r.Context(), accountID, params, body, gw.bedrockResolver(), gw.bedrockEndpointResolver())
 	if err != nil {
 		return err
diff --git a/tests/e2e/bedrock/bedrock_test.go b/tests/e2e/bedrock/bedrock_test.go
index 15d57b7e1..ffee8260b 100644
--- a/tests/e2e/bedrock/bedrock_test.go
+++ b/tests/e2e/bedrock/bedrock_test.go
@@ -3,6 +3,7 @@
 package bedrock
 
 import (
+	"encoding/json"
 	"os"
 	"testing"
 
@@ -118,6 +119,77 @@ func TestConverseAnthropic(t *testing.T) {
 	assert.Greater(t, aws.Int64Value(out.Usage.OutputTokens), int64(0), "OutputTokens must be > 0")
 }
 
+// TestInvokeModelSelfHost exercises the real vLLM InvokeModel path with the
+// Bedrock-native Llama request/response shape. Gated behind
+// OCHRE_E2E_SELFHOST=1 — it needs a GPU-backed vLLM endpoint configured on the
+// gateway via OCHRE_VLLM_ENDPOINTS at boot.
+func TestInvokeModelSelfHost(t *testing.T) {
+	if os.Getenv("OCHRE_E2E_SELFHOST") != "1" {
+		t.Skip("OCHRE_E2E_SELFHOST!=1; skipping real self-host InvokeModel")
+	}
+	f := requireBedrockFixture(t)
+
+	reqBody, err := json.Marshal(map[string]any{
+		"prompt":      "Say hello in one word.",
+		"max_gen_len": 64,
+	})
+	require.NoError(t, err, "marshal invoke-model request")
+
+	out, err := f.AWS.BedrockRuntime.InvokeModel(&bedrockruntime.InvokeModelInput{
+		ModelId:     aws.String(modelSelfHostLlama),
+		Body:        reqBody,
+		ContentType: aws.String("application/json"),
+	})
+	require.NoError(t, err, "invoke-model self-host")
+	assert.Equal(t, "application/json", aws.StringValue(out.ContentType))
+
+	var respBody struct {
+		Generation           string `json:"generation"`
+		GenerationTokenCount int    `json:"generation_token_count"`
+	}
+	require.NoError(t, json.Unmarshal(out.Body, &respBody), "unmarshal invoke-model response")
+	assert.NotEmpty(t, respBody.Generation, "empty generation")
+	assert.Greater(t, respBody.GenerationTokenCount, 0, "GenerationTokenCount must be > 0")
+}
+
+// TestInvokeModelAnthropic exercises the real Anthropic-direct InvokeModel
+// path with the Bedrock-native Claude Messages request/response shape. Gated
+// behind OCHRE_E2E_ANTHROPIC=1 — it needs a live Anthropic API key configured
+// on the cluster (OCHRE_ANTHROPIC_API_KEY or a per-account KV entry).
+func TestInvokeModelAnthropic(t *testing.T) {
+	if os.Getenv("OCHRE_E2E_ANTHROPIC") != "1" {
+		t.Skip("OCHRE_E2E_ANTHROPIC!=1; skipping real Anthropic InvokeModel")
+	}
+	f := requireBedrockFixture(t)
+
+	reqBody, err := json.Marshal(map[string]any{
+		"anthropic_version": "bedrock-2023-05-31",
+		"max_tokens":        64,
+		"messages": []map[string]any{
+			{"role": "user", "content": "Say hello in one word."},
+		},
+	})
+	require.NoError(t, err, "marshal invoke-model request")
+
+	out, err := f.AWS.BedrockRuntime.InvokeModel(&bedrockruntime.InvokeModelInput{
+		ModelId:     aws.String(modelAnthropic),
+		Body:        reqBody,
+		ContentType: aws.String("application/json"),
+	})
+	require.NoError(t, err, "invoke-model anthropic")
+	assert.Equal(t, "application/json", aws.StringValue(out.ContentType))
+
+	var respBody struct {
+		Content []struct {
+			Type string `json:"type"`
+			Text string `json:"text"`
+		} `json:"content"`
+	}
+	require.NoError(t, json.Unmarshal(out.Body, &respBody), "unmarshal invoke-model response")
+	require.NotEmpty(t, respBody.Content, "empty content")
+	assert.NotEmpty(t, respBody.Content[0].Text, "empty assistant text")
+}
+
 // converseText concatenates the text content blocks of a Converse output message.
 func converseText(msg *bedrockruntime.Message) string {
 	var text string