cerebras

/

Qwen3-Coder-REAP-25B-A3B

like 37

Follow

Cerebras 990

Here are the results of a single pass through the current Aider benchmark. These results are for the non-quant'd as per this repo.

pass_rate_2 seems to be a little lower than the reported rate for the default BF16. In the Aider Discord, they seem to report BF16 to be 32.4.

I'm a newb when it comes to Aider and vllm, so I've also included the command lines I used in case there is something wrong.

Serve:

CUDA_VISIBLE_DEVICES=0,1 \
  vllm serve \
  --port 5000 \
  --gpu-memory-utilization=0.95 \
  --max_model_len=131072 \
  --pipeline-parallel-size 2 \
  --model=/home/ai/models/cerebras/Qwen3-Coder-REAP-25B-A3B \
  --served-model-name Qwen3-Coder-REAP-25B-A3B

Aider:

./benchmark/benchmark.py Qwen3-Coder-REAP-25B-A3B --model openai/Qwen3-Coder-REAP-25B-A3B --edit-format diff --threads 1 --exercises-dir polyglot-benchmark

Results:

- dirname: 2025-10-22-11-21-46--Qwen3-Coder-REAP-25B-A3B
  test_cases: 225
  model: openai/Qwen3-Coder-REAP-25B-A3B
  edit_format: diff
  commit_hash: 11516d6
  pass_rate_1: 12.0
  pass_rate_2: 27.6
  pass_num_1: 27
  pass_num_2: 62
  percent_cases_well_formed: 94.7
  error_outputs: 20
  num_malformed_responses: 19
  num_with_malformed_responses: 12
  user_asks: 114
  lazy_comments: 1
  syntax_errors: 0
  indentation_errors: 0
  exhausted_context_windows: 1
  prompt_tokens: 3341547
  completion_tokens: 549320
  test_timeouts: 5
  total_tests: 225
  command: aider --model openai/Qwen3-Coder-REAP-25B-A3B
  date: 2025-10-22
  versions: 0.86.2.dev
  seconds_per_case: 35.2
  total_cost: 0.0000