diff --git a/src/chap/backends/llama_cpp.py b/src/chap/backends/llama_cpp.py
index 381b825..99ddece 100644
--- a/src/chap/backends/llama_cpp.py
+++ b/src/chap/backends/llama_cpp.py
@@ -18,10 +18,10 @@ class LlamaCpp(AutoAskMixin):
         url: str = "http://localhost:8080/completion"
         """The URL of a llama.cpp server's completion endpoint."""
 
-        start_prompt: str = """<s>[INST] <<SYS>>\n"""
-        after_system: str = "\n<</SYS>>\n\n"
-        after_user: str = """ [/INST] """
-        after_assistant: str = """ </s><s>[INST] """
+        start_prompt: str = "<s>"
+        system_format: str = "<<SYS>>{}<</SYS>>"
+        user_format: str = " [INST] {} [/INST]"
+        assistant_format: str = " {}</s>"
 
     def __init__(self) -> None:
         super().__init__()
@@ -34,18 +34,18 @@ A dialog, where USER interacts with AI. AI is helpful, kind, obedient, honest, a
     def make_full_query(self, messages: Session, max_query_size: int) -> str:
         del messages[1:-max_query_size]
         result = [self.parameters.start_prompt]
+        formats = {
+            Role.SYSTEM: self.parameters.system_format,
+            Role.USER: self.parameters.user_format,
+            Role.ASSISTANT: self.parameters.assistant_format,
+        }
         for m in messages:
             content = (m.content or "").strip()
             if not content:
                 continue
-            result.append(content)
-            if m.role == Role.SYSTEM:
-                result.append(self.parameters.after_system)
-            elif m.role == Role.ASSISTANT:
-                result.append(self.parameters.after_assistant)
-            elif m.role == Role.USER:
-                result.append(self.parameters.after_user)
+            result.append(formats[m.role].format(content))
         full_query = "".join(result)
+        print("fq", full_query)
         return full_query
 
     async def aask(
diff --git a/src/chap/backends/mistral.py b/src/chap/backends/mistral.py
new file mode 100644
index 0000000..c254632
--- /dev/null
+++ b/src/chap/backends/mistral.py
@@ -0,0 +1,99 @@
+# SPDX-FileCopyrightText: 2024 Jeff Epler <jepler@gmail.com>
+#
+# SPDX-License-Identifier: MIT
+
+import json
+from dataclasses import dataclass
+from typing import AsyncGenerator, Any
+
+import httpx
+
+from ..core import AutoAskMixin
+from ..key import get_key
+from ..session import Assistant, Session, User
+
+
+class Mistral(AutoAskMixin):
+    @dataclass
+    class Parameters:
+        url: str = "https://api.mistral.ai"
+        model: str = "open-mistral-7b"
+        max_new_tokens: int = 1000
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.parameters = self.Parameters()
+
+    system_message = """\
+Answer each question accurately and thoroughly.
+"""
+
+    def make_full_query(self, messages: Session, max_query_size: int) -> dict[str, Any]:
+        messages = [m for m in messages if m.content]
+        del messages[1:-max_query_size]
+        result = dict(
+            model=self.parameters.model,
+            max_tokens=self.parameters.max_new_tokens,
+            messages=[dict(role=str(m.role), content=m.content) for m in messages],
+            stream=True,
+        )
+        return result
+
+    async def aask(
+        self,
+        session: Session,
+        query: str,
+        *,
+        max_query_size: int = 5,
+        timeout: float = 180,
+    ) -> AsyncGenerator[str, None]:
+        new_content: list[str] = []
+        params = self.make_full_query(session + [User(query)], max_query_size)
+        try:
+            async with httpx.AsyncClient(timeout=timeout) as client:
+                async with client.stream(
+                    "POST",
+                    f"{self.parameters.url}/v1/chat/completions",
+                    json=params,
+                    headers={
+                        "Authorization": f"Bearer {self.get_key()}",
+                        "content-type": "application/json",
+                        "accept": "application/json",
+                        "model": "application/json",
+                    },
+                ) as response:
+                    if response.status_code == 200:
+                        async for line in response.aiter_lines():
+                            if line.startswith("data:"):
+                                data = line.removeprefix("data:").strip()
+                                if data == "[DONE]":
+                                    break
+                                j = json.loads(data)
+                                content = (
+                                    j.get("choices", [{}])[0]
+                                    .get("delta", {})
+                                    .get("content", "")
+                                )
+                                if content:
+                                    new_content.append(content)
+                                    yield content
+                    else:
+                        content = f"\nFailed with {response=!r}"
+                        new_content.append(content)
+                        yield content
+                        async for line in response.aiter_lines():
+                            new_content.append(line)
+                            yield line
+        except httpx.HTTPError as e:
+            content = f"\nException: {e!r}"
+            new_content.append(content)
+            yield content
+
+        session.extend([User(query), Assistant("".join(new_content))])
+
+    @classmethod
+    def get_key(cls) -> str:
+        return get_key("mistral_api_key")
+
+
+factory = Mistral