DeepSeek/Qwen Structured Calls Succeed but Return Empty? Disable Thinking Before It Burns Your Token Budget
While running batch LLM semantic validation over production data, all 2,449 calls "succeeded" โ yet every single result fell back to the default value, and the logs contained almost no failures.
Encountered this while building AI Ops โ LLM-powered analysis that surfaces market trends, user behavior, and sales insights to drive precise operations strategy. The product title optimization pipeline asks an LLM to semantically validate "keyword ร product" pairs: one call per keyword batch, returning a tiny JSON judgment. This should be the simplest kind of LLM call โ yet on the first production run, all 2,449 pairs silently fell back, and the validation layer produced zero effective LLM judgments.
TL;DRโ
Thinking-capable models like DeepSeek and Qwen share a single max_tokens budget between reasoning text and final content. If structured small-output calls don't explicitly disable thinking, the reasoning chain eats the entire budget on its own: finish_reason becomes length, content comes back empty โ and since the parse-retry loop only logs exceptions, the silent retries exhaust and degrade the whole batch without a single error. Two things to do: disable thinking explicitly for structured calls, and log finish_reason on parse failure instead of only catching exceptions.
The Symptomโ
This minimal reproduction shows the whole process (requires pip install openai and a thinking-capable model):
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["DEEPSEEK_API_KEY"],
base_url="https://api.deepseek.com",
)
resp = client.chat.completions.create(
model="deepseek-reasoner",
messages=[
{
"role": "user",
"content": (
'ๅคๆญไธ้ข็ๅ
ณ้ฎ่ฏๆฏๅฆ้ๅๅๅ
ฅๅๅๆ ้ข๏ผ'
'ๅช่ฟๅ JSON๏ผ{"suitable": true} ๆ {"suitable": false}ใ\n'
"ๅ
ณ้ฎ่ฏ๏ผsummer women dress\n"
"ๅๅ๏ผfloral midi dress for women"
),
}
],
max_tokens=2000, # reasoning and content share this budget
)
print("finish_reason:", resp.choices[0].finish_reason)
print("content:", repr(resp.choices[0].message.content))
Typical output with thinking enabled:
finish_reason: length
content: ''
Everything looks fine at the HTTP layer: no timeout, no 5xx, no SDK exception. If the outer layer is a "retry on parse failure, fall back after retries exhaust" loop, the logs end up with a single "retries exhausted" warning โ which reads exactly like an intermittent network problem.
Root Causeโ
Thinking models don't budget reasoning separately. In DeepSeek, Qwen and similar models, reasoning content and final content share the same max_tokens ceiling โ there is no independent "reasoning budget" field.
Structured small-output calls tend to use small budgets. A boolean judgment is expected to produce a few dozen tokens, so max_tokens=2000 looks generous โ but the reasoning chain's length is completely uncontrolled. Once it consumes all 2,000 tokens, the model never gets a chance to emit the actual answer: finish_reason returns length, content is an empty string, yet the API returns 200 normally.
There's also an amplifier on the engineering side. An empty string is not valid JSON, but many retry loops only catch network and API exceptions, treating parse failure as "no result this round" and retrying silently. This kind of exception-swallowing silent failure is notoriously hard to diagnose inside retry loops: all N retries fail for the same root cause, yet the log shows only the final fallback warning โ easy to misread as network flakiness.
The Fixโ
Step 1: Explicitly disable thinking for structured-output callsโ
Boolean judgments, JSON extraction, and classification calls don't need multi-step reasoning. Disable thinking per provider:
def thinking_disabled_extra_body(provider: str) -> dict:
"""Structured small-output calls: disable thinking explicitly per provider."""
if provider == "deepseek":
return {"thinking": {"type": "disabled"}}
if provider == "qwen":
return {"enable_thinking": False}
return {}
resp = client.chat.completions.create(
model=model_name,
messages=messages,
max_tokens=2000,
extra_body=thinking_disabled_extra_body("deepseek"),
)
With thinking off, the entire 2,000-token budget goes to the JSON judgment itself. After the fix, rerunning the batch returned valid judgments for all 2,449 keywordรproduct pairs โ before the fix, the whole batch silently degraded, leaving just 3 "retries exhausted" warnings in the logs.
Step 2: Don't let parse failures go silentโ
Even with thinking disabled, turn "parse failure" into an evidence-bearing log line, so next time an empty output occurs โ for any reason โ a single log line locates it:
import json
import logging
logger = logging.getLogger(__name__)
def parse_judgment(resp) -> dict | None:
content = resp.choices[0].message.content
try:
return json.loads(content)
except (TypeError, json.JSONDecodeError):
# Key: log finish_reason and raw content, not just exceptions
logger.warning(
"LLM output parse failed: finish_reason=%s content=%r",
resp.choices[0].finish_reason,
content,
)
return None
When debugging LLM degradation, check finish_reason first: length means the output budget was exhausted (most likely by reasoning), while stop means normal completion. This is far more effective than scrolling exception logs.
If your returned JSON also passes through schema validation and you use Zod in a TypeScript project, watch out for this related pitfall: Zod schema validation silently dropping LLM output.
Heads up
The parameter to disable thinking is not standardized across providers: DeepSeek uses {"thinking": {"type": "disabled"}}, Qwen uses {"enable_thinking": False}, and OpenAI's o-series uses reasoning_effort-style parameters. Check each provider's docs before integrating โ don't assume the parameter is universal.
If a provider's thinking cannot be disabled, you must raise max_tokens based on measured reasoning length โ otherwise the same silent degradation will happen again.
FAQโ
How do I disable thinking output in DeepSeek?โ
On the OpenAI-compatible API, pass {"thinking": {"type": "disabled"}} via extra_body; Qwen uses {"enable_thinking": False}. Structured small-output calls like boolean judgments and JSON extraction don't need a reasoning chain โ turn it off by default so the entire budget goes to the actual answer.
Why does my DeepSeek API call succeed but return empty content?โ
Thinking models share max_tokens between reasoning and content. When reasoning exhausts the budget, finish_reason returns length and content is empty โ with no exception thrown. Disable thinking or raise max_tokens based on measured reasoning length, and check finish_reason before digging through exception logs.
CCLEE
Independent developer, 24 years in e-commerce, focused on grounding AI in real business scenarios.
Work with me