llm.create_chat_completion(
messages = [
{
"role": "user",
"content": "What is the capital of France?"
}
]
)Qwen3-Chimera-60B-A3B-Q4_K_M
Experimental MoE expert-bank fusion model combining Qwen3-30B-Instruct and Qwen3-Coder-30B for local inference.
Model Summary Qwen3-Chimera-60B-A3B-Q4_K_M is an experimental fused MoE model built from the expert banks of:
qwen3:30b-instruct qwen3-coder:30b It is designed to preserve general chat behavior, coding behavior, and basic arithmetic / reasoning utility while keeping the active path small enough for local inference.
This is not a newly trained foundation model. It is an unofficial expert-bank fusion experiment.
Key Specs Formal name: Qwen3-Chimera-60B-A3B-Q4_K_M Architecture: qwen3moe Total parameters: 59.536B Rounded public name: 60B Quantization: Q4_K_M Model file size: 36,160,180,224 bytes Size on disk: 36.16 GB / 33.68 GiB Context length: 262144 Expert count: 256 Routing: top-8 routed experts per token Estimated active parameters: ~`3.0B` What Makes It Different This model was built to combine the strongest practical traits of both source models:
Qwen3-30B-Instruct contributes general chat behavior, broad instruction following, and language coverage. Qwen3-Coder-30B contributes coding-oriented behavior, structured output, and engineering-style responses. The result is a fused model that keeps useful general behavior while remaining practical for local inference.
Technical Challenges Addressed
Prompt template alignment The first important fix was aligning the deployment chat template to the original qwen3:30b-instruct role-based format. Using a bare {{ .Prompt }} wrapper caused short chat prompts to behave inconsistently and made the model look much worse than it actually was.
Chinese prompt correctness Some early tests were affected by mojibake / wrong prompt encoding paths. Once the model was tested through the proper api/chat path with UTF-8 content, Traditional Chinese responses became normal.
Top-k stability Several top-k variants were tested:
top6 was fast but unstable on short chat top8 became the best public-facing tradeoff top10 and top12 were usable, but slower 4. Local inference constraints This model is intended for local use. The design goal was not to create a new foundation model from scratch, but to preserve useful behavior while keeping active compute low.
Observed Behavior In internal testing, the model:
answers in Traditional Chinese normally through api/chat handles short normal chat well once the correct template is used produces usable Python/code answers handles simple arithmetic and short reasoning prompts gives respectful refusals for inappropriate subjective comparisons involving public figures Internal Benchmark Notes These are small internal checks, not a formal leaderboard submission:
Code prompts: matched qwen3:30b-instruct on a short 5-prompt set Math prompts: matched qwen3-coder:30b on a short 5-prompt set Chinese chat: normal and usable through the proper chat interface Recommended Use Good for:
local chat Chinese chat code assistance short reasoning arithmetic mixed general-purpose local use Not ideal for:
claims of being a newly trained 60B foundation model claims of full benchmark dominance over both source models use cases that require a fully verified, retrained, official model release Provenance This is a fusion experiment based on the expert banks of two upstream Qwen3 models.
Notes This model is experimental and unofficial. The 60B label is a rounded public name. The measured total parameter count is 59.536B. The active path at top-8 is estimated at about 3.0B parameters.
- Downloads last month
- 42
4-bit
Model tree for XxACCOxX/qwen3-chimera-60b-a3b-q4-k-m
Base model
Qwen/Qwen3-30B-A3B-Instruct-2507
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="XxACCOxX/qwen3-chimera-60b-a3b-q4-k-m", filename="Qwen3-Chimera-60B-A3B-Q4_K_M.gguf", )