Morph

Use Morph through LiteLLM for open-weight chat models and Fast Apply code edits. Set MORPH_API_KEY and prefix the model ID with morph/. LiteLLM uses https://api.morphllm.com/v1 by default.

Quickstart

Create a key in the Morph dashboard, then set it in your environment.

pip install -U 'litellm[proxy]'
export MORPH_API_KEY="<your-morph-api-key>"

from litellm import completion

response = completion(
    model="morph/morph-dsv41flash",
    messages=[{"role": "user", "content": "Write a Python function that adds two numbers."}],
)
print(response.choices[0].message.content)

Open-weight models

Morph documents a 1M-token context window for these models. Use the exact LiteLLM model string in SDK calls or proxy configuration.

Model	LiteLLM model string
Kimi K3	morph/morph-kimik3
Kimi K3 Fast	morph/morph-kimik3-fast
GLM-5.3	morph/morph-glm53-744b
GLM-5.3-Flash	morph/morph-glm53flash
DeepSeek V4.1 Flash	morph/morph-dsv41flash
DeepSeek V4.1 Flash 1017	morph/morph-dsv41flash-1017
DeepSeek V4.1 Flash Abliterated	morph/morph-dsv41flash-abliterated
DeepSeek V4 Flash 0731	morph/morph-dsv4flash

The abliterated model requires approval from Morph. Model access also depends on your account quota and available provider capacity.

Open-weight models accept tools, response_format, and reasoning_effort. DeepSeek V4.1 Flash enables reasoning by default. Use reasoning_effort="none" to disable it, or omit the parameter to use Morph's default.

Use the AI Gateway

Add this deployment to config.yaml.

model_list:
  - model_name: morph-deepseek
    litellm_params:
      model: morph/morph-dsv41flash
      api_key: os.environ/MORPH_API_KEY

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY

Generate a gateway key, then start the gateway with both keys exported in the same environment. The Morph key authenticates the upstream request; the LiteLLM key authenticates your application to the gateway.

export LITELLM_MASTER_KEY="sk-$(openssl rand -hex 32)"
litellm --config config.yaml --port 4000

Send a request to the local gateway. The response includes the model's answer in choices[0].message.content and token counts in usage.

curl http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "morph-deepseek",
    "messages": [{"role": "user", "content": "Reply with exactly OK."}],
    "reasoning_effort": "none"
  }'

For a shared deployment, give applications virtual keys instead of the gateway's master key.

Streaming and prefix caching

Morph caches matching prompt prefixes automatically. Keep the system prompt, tool definitions, and existing messages unchanged between turns. Send a unique x-session-id for each conversation to improve worker affinity.

from uuid import uuid4
from litellm import completion

session_id = str(uuid4())

response = completion(
    model="morph/morph-dsv41flash",
    messages=[{"role": "user", "content": "Explain Python generators."}],
    extra_headers={"x-session-id": session_id},
    stream=True,
    stream_options={"include_usage": True},
)

for chunk in response:
    if chunk.choices:
        print(chunk.choices[0].delta.content or "", end="")
    if chunk.usage:
        print(chunk.usage)

Reuse session_id for later turns in that conversation. Do not share one value across unrelated conversations. prompt_cache_key is also supported, but Morph documents it as a region-placement hint, while x-session-id selects a worker.

usage.prompt_tokens includes both cached and uncached input. usage.prompt_tokens_details.cached_tokens counts reused tokens. LiteLLM calculates cost from the model's catalog rates:

uncached input = prompt_tokens - cached_tokens
cost = uncached input * input rate
     + cached_tokens * cache-read rate
     + completion_tokens * output rate

Morph charges new prefixes at the normal input rate, with no cache-write surcharge. It does not report a separate cache-creation token count on the Chat Completions responses tested here. LiteLLM does not infer that every uncached token was stored. Reasoning tokens count toward billed output.

See Morph's pricing page and caching guide. The catalog uses the direct pricing page and caching guide. Morph's machine-readable catalog currently lists different rates for several models. Treat LiteLLM's calculated cost as an estimate until Morph reconciles those sources, and verify charges against your Morph account.

Tools and structured output

Use the OpenAI tool format for tools and tool_choice. For a strict response shape, pass a JSON schema.

from litellm import completion

response = completion(
    model="morph/morph-dsv41flash",
    messages=[{"role": "user", "content": "Return JSON with status equal to ok."}],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "status",
            "strict": True,
            "schema": {
                "type": "object",
                "properties": {"status": {"type": "string", "enum": ["ok"]}},
                "required": ["status"],
                "additionalProperties": False,
            },
        },
    },
)
print(response.choices[0].message.content)

Fast Apply

morph/morph-v3-fast and morph/morph-v3-large apply edits to existing code. Give them an instruction, the original code, and the proposed update. They use a different prompt from the chat models and do not support the chat-only options above.

from litellm import completion

response = completion(
    model="morph/morph-v3-fast",
    messages=[{
        "role": "user",
        "content": """<instruction>Change the greeting to Hello, world!</instruction>
<code>print("Hello")</code>
<update>print("Hello, world!")</update>""",
    }],
)
print(response.choices[0].message.content)

Use stream=True to stream the edit. See Morph's Fast Apply guide for prompt requirements.

Endpoint and error handling

Override the API base with api_base or MORPH_API_BASE. Pass an explicit key with api_key when you do not use the environment variable.

A 401 indicates an authentication problem. A 402 with Monthly quota exceeded requires more Morph quota. A 429 can indicate a rate limit or unavailable model capacity; respect the provider's retry guidance.