Max Tokens do not work at all for Openrouter

[jfangzr]

What happened?

I expected that the entire context window would be given to me, but apparently not at all. It seems like the window is cutoff at 2048 or 4096 tokens. I was using qwen/qwen3-235b-a22b-thinking-2507 via OpenRouter, and I specifically set the max_tokens parameter to 131072. Well, it did not work. I also tried in the UI to set the preset to 100000 for both max context tokens and max tokens, but this is still not working. The output is still truncated after just a few thousand tokens at most. I am following the instructions exactly, to the point, so I don't know what is going wrong at all. I don't know why on earth there is token number cutoff, I thought librechat did not impose any output token limits itself?

also, I realized I used the wrong github account to create the previous issue, I am going to close the previous issue #9682

Version Information

$ docker images | grep librechat
ghcr.io/danny-avila/librechat-dev latest d7aeafdec1ac 5 days ago 1.57GB
ghcr.io/danny-avila/librechat-rag-api-dev-lite latest d813432c9b0a 3 weeks ago 1.68GB

Steps to Reproduce

1. set max_tokens parameter to 131072 under addParams under openrouter in librechat.yaml
2. use openrouter, specifically, model provider cerebras, with model set as wen/qwen3-235b-a22b-thinking-2507
3. put in the following prompt: The 9 members of a baseball team went to an ice-cream parlor after their game. Each player had a single scoop cone of chocolate, vanilla, or strawberry ice cream. At least one player chose each flavor, and the number of players who chose chocolate was greater than the number of players who chose vanilla, which was greater than the number of players who chose strawberry. Let N be the number of different assignments of flavors to players that meet these conditions. Find the remainder when N is divided by 1000.
4. rule out other causes; the free api key works very well for cerebras, and calling the openrouter api key from python also works correctly. somehow calling it from librechat, however, the truncation happens.

What browsers are you seeing the problem on?

No response

Relevant log output

2025-09-18T03:04:45.805Z debug: [spendTokens] conversationId: 4ec43627-be1c-49e8-b55b-35d76ae47c98 | Context: message | Token usage:
{
  promptTokens: 12,
  completionTokens: 565,
}
2025-09-18T03:04:45.826Z debug: [spendTokens] No transactions incurred against balance
2025-09-18T03:10:06.271Z debug: [spendTokens] conversationId: 007b336b-a307-49a4-b12e-b196e91390b8 | Context: message | Token usage:
{
  promptTokens: 16,
  completionTokens: 1597,
}
2025-09-18T03:10:06.283Z debug: [spendTokens] No transactions incurred against balance
2025-09-18T03:10:08.412Z debug: [spendTokens] conversationId: 007b336b-a307-49a4-b12e-b196e91390b8 | Context: title | Token usage:
{
  promptTokens: 1137,
  completionTokens: 591,
}
2025-09-18T03:10:08.421Z debug: [spendTokens] No transactions incurred against balance
2025-09-18T03:22:17.363Z debug: [spendTokens] conversationId: a15943b1-d495-406c-a093-12d902c025a5 | Context: message | Token usage:
{
  promptTokens: 123,
  completionTokens: 4096,
}
2025-09-18T03:22:17.399Z debug: [spendTokens] No transactions incurred against balance
2025-09-18T03:22:20.847Z debug: [spendTokens] conversationId: a15943b1-d495-406c-a093-12d902c025a5 | Context: title | Token usage:
{
  promptTokens: 154,
  completionTokens: 1771,
}
2025-09-18T03:22:20.856Z debug: [spendTokens] No transactions incurred against balance

Screenshots

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

First off, if the system correctly fetches the max context values from OpenRouter, it will have the expected amount without needing to set anything in the UI. You should check your error logs if the system is having trouble fetching the official model list from OR, which includes context window values:

2025-09-18T12:26:32.264Z debug: [BaseClient] Context Count (1/2)
{
  remainingContextTokens: 235921,
  maxContextTokens: 235930,
}
2025-09-18T12:26:32.264Z debug: [BaseClient] Context Count (2/2)
{
  remainingContextTokens: 235921,
  maxContextTokens: 235930,
}
2025-09-18T12:26:32.264Z debug: [BaseClient] tokenCountMap:
{
  83b5591e-ea6d-431a-9078-e91311aef796: 6,
}
2025-09-18T12:26:32.264Z debug: [BaseClient]
{
  promptTokens: 9,
  remainingContextTokens: 235921,
  payloadSize: 1,
  maxContextTokens: 235930,
}
2025-09-18T12:26:32.264Z debug: [api/server/controllers/agents/client.js #useMemory] User 682f49b90f07376815c38ef2 does not have USE permission for memories
2025-09-18T12:26:32.264Z debug: [BaseClient] tokenCountMap
{
  83b5591e-ea6d-431a-9078-e91311aef796: 6,
}
2025-09-18T12:26:32.264Z debug: [BaseClient] userMessage
{
  messageId: "83b5591e-ea6d-431a-9078-e91311aef796",
  parentMessageId: "00000000-0000-0000-0000-000000000000",
  conversationId: "4587b6f4-42bd-4eb9-8cd7-db2fba0eb94e",
  sender: "User",
  text: "hi there",
  isCreatedByUser: true,
  tokenCount: 6,
}
2025-09-18T12:26:32.281Z debug: [Balance.check] Initial state
{
  user: "some-user-id",
  model: "qwen/qwen3-235b-a22b-thinking-2507",
  endpoint: "OpenRouter",
  valueKey: undefined,
  tokenType: "prompt",
  amount: 9,
  balance: 24953073.24000012,
  multiplier: 6,
  endpointTokenConfig: false,
}

and I specifically set the max_tokens parameter to 131072

I don't know where you set this parameter, but max_tokens is not for the window but for the response. In other words, this is to control the maximum output tokens.

What you want is here:

After setting this 33111 amount, it's correctly being used:

2025-09-18T12:30:51.919Z debug: [BaseClient] Context Count (1/2)
{
  remainingContextTokens: 29551,
  maxContextTokens: 29800,
}
2025-09-18T12:30:51.920Z debug: [BaseClient] Context Count (2/2)
{
  remainingContextTokens: 29551,
  maxContextTokens: 29800,
}
2025-09-18T12:30:51.920Z debug: [BaseClient] tokenCountMap:
{
  83b5591e-ea6d-431a-9078-e91311aef796: 11,
  bd9ed3ee-cdc3-49c4-ad2d-64255f2d1872: 229,
  688e4d6c-98b1-471c-b668-a905b5491ec3: 6,
}
2025-09-18T12:30:51.920Z debug: [BaseClient]
{
  promptTokens: 249,
  remainingContextTokens: 29551,
  payloadSize: 3,
  maxContextTokens: 29800,
}

29,800 is roughly ~90% of 33,111, as we will conservatively always use 90% of the configured/system context window.

[jfangzr]

Truncation is still happening.

I have also tried to remove the

addParams:
  max_tokens: 65536

part from LibreChat.yaml and it is still not working correctly, still truncating.

Also, when I typed

cat error-2025-09-18.log

and also read the error logs for 2025-09-17, both of them show empty. So I don't think there was a processing error at all. This would suggest that there is a bug somewhere...

[danny-avila]

Don't utilize addParams for this. Also, you will have issues if max output tokens are larger than max context tokens.

You will want to set these amounts via an agent, preset, or model spec:

You can even use this URL to help pre-populate values:

http://localhost:3080/c/new?endpoint=OpenRouter&model=qwen%2Fqwen3-235b-a22b-thinking-2507&maxContextTokens=262000&max_tokens=131000

According to OR, the max context is 262k, and max output varies by provider, from 131K to the context limit.

https://openrouter.ai/qwen/qwen3-235b-a22b-thinking-2507

[danny-avila]

set the preset to 100000 for both max context tokens and max tokens

If you set these both to the same value or max tokens larger than context, you will have "truncation" issues.

In general, you always want lower than the max context for output, usually anywhere from 8k-32k depending on the use case.