GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
bayliner1980
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
bayliner1980
Model Tree
Qwen/Qwen3-0.6B
Input Modalities
Output Modalities
Supported Functionality
| Qwen3-0.6B (dense) | This model | |
|---|---|---|
| Architecture | Qwen3ForCausalLM | Qwen3MoeForCausalLM |
| Layers / hidden size / heads (q, kv) | 28 / 1024 / 16, 8 | same |
| MLP per layer | 1 x 3072 | 32 experts x 96, top-16, norm_topk_prob |
| Total parameters | 596.0M | 597.0M (+0.9M router weights) |
| Active parameters per token | 596.0M | 464.9M (309.3M excluding embeddings) |
| Active MLP compute | 100% | 50% |
| Context length used in training / eval | 512 / 1024 tokens |
Attention, embeddings, norms and the (tied) LM head are the original Qwen3-0.6B weights, unchanged. The same 3072 MLP neurons per layer are regrouped into the 32 experts, so the total parameter count barely changes; the saving is in compute per token.
Perplexity: wikitext-2 test (1024-token chunks) and 64 x 1024 tokens of C4 validation. Tasks: zero-shot acc_norm,
1000 items per task (500 for OpenBookQA), scored with this project's own harness by per-character normalised
log-likelihood. These are not lm-evaluation-harness numbers, so compare them only with each other.
| Qwen3-0.6B (dense) | This model | |
|---|---|---|
| Wikitext-2 ppl ↓ | 23.88 | 26.98 (+13.0%) |
| C4 ppl ↓ | 31.63 | 36.32 (+14.8%) |
| HellaSwag | 0.480 | 0.458 |
| ARC-Easy | 0.578 | 0.503 |
| ARC-Challenge | 0.339 | 0.283 |
| WinoGrande | 0.571 | 0.558 |
| OpenBookQA | 0.308 | 0.322 |
| Task mean | 0.455 | 0.425 |
On the tasks it keeps about 81% of the dense model's above-chance accuracy. The exported model reproduces the
converted checkpoint (wikitext-2 26.978 vs 26.985). The bf16 GGUF matches transformers to within 0.08%; Q4_K_M
adds +4.1% perplexity (the dense Qwen3-0.6B loses 7.9% at Q4_K_M).
Chat format works: on short chat exchanges it predicts the <|im_end|> that ends its turn with probability 0.75
(dense: 0.79) and the empty <think></think> block of non-thinking mode as the dense model does, and it ends its
turn by itself in greedy generation.
Greedy decoding, same prompts for both models; chat with enable_thinking=False. Picked to show both what works and
what doesn't.
Chat: "What is the capital of Australia?"
| Qwen3-0.6B (dense) | This model |
|---|---|
| The capital of Australia is Canberra. | The capital of Australia is Canberra. |
Chat: "What is 15% of 240?"
| Qwen3-0.6B (dense) | This model |
|---|---|
| Converts 15% to 0.15, multiplies, and concludes: So, 15% of 240 is 36. | 15% of 240 is 36. |
Chat: "Give me three tips for sleeping better."
| Qwen3-0.6B (dense) | This model |
|---|---|
| 1. Avoid screens before bed … 2. Create a relaxing bedtime routine … 3. Limit caffeine and alcohol … | 1. Immerse yourself in a relaxing environment – Try a warm bath … 2. Limit screen time before bed – Avoid devices that emit blue light … 3. Create a consistent bedtime routine – Establish a regular sleep schedule … |
Chat: "Who painted the Mona Lisa, and when?"
| Qwen3-0.6B (dense) | This model |
|---|---|
| The Mona Lisa is a famous painting by Leonardo da Vinci, and it was painted between 1503 and 1506. | The Mona Lisa was painted by Leonardo da Vinci in 1519. (wrong year) |
Chat: "Explain why the sky is blue in two sentences."
| Qwen3-0.6B (dense) | This model |
|---|---|
| The sky appears blue because of the way light interacts with the Earth's atmosphere. Short wavelengths of light, such as blue, are scattered more efficiently by the atmosphere than longer wavelengths like red or orange, making the sky appear predominantly blue. | The sky is blue because it contains a mixture of various colors, primarily blue, which is the result of the combination of different pigments in the atmosphere. (wrong) |
Completion: "Once upon a time, in a small village by the sea,"
| Qwen3-0.6B (dense) | This model |
|---|---|
| there was a young girl named Lila. She was known for her kindness and her ability to help others. One day, she decided to start a small business that helped the poor and the elderly. … | there was a boy named Sam who loved to play with a small toy boat. The boat was made of a special material, and it could float on water. The village was on the edge of a lagoon … |
Completion: Python
python
def fibonacci(n): """Return the n-th Fibonacci number."""Qwen3-0.6B (dense):
python
if n == 0: return 0 elif n == 1: return 1 else: return fibonacci(n-1) + fibonacci(n-2)This model:
python
for n in range(n - 1, n + 1): return n - 1 + 1 return nIn short: fluent text and chat formatting close to the dense model, and common facts and simple arithmetic usually right. Specific facts (dates, mechanisms) are often confidently wrong, and code is broken; the distillation data was educational web text with almost no code.
python
from transformers import AutoModelForCausalLM, AutoTokenizer repo = "bayliner1980/Qwen3-0.6B-MoE-A0.5B"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto") messages = [{"role": "user", "content": "Write one sentence about the sea."}]ids = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_tensors="pt", return_dict=True).to(model.device)out = model.generate(**ids, max_new_tokens=128, do_sample=True, temperature=0.7, top_p=0.8)print(tok.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))It was exported and tested with transformers 5.5.0. The config stores the expert count as num_local_experts;
older transformers versions expect num_experts and will not load it correctly. Use it with thinking off
(enable_thinking=False); thinking mode was never trained after conversion.
enable_thinking=False.If you find this work useful or interesting, any contribution helps.
Apache 2.0, the same as the base model Qwen/Qwen3-0.6B.
def fibonacci(n): """Return the n-th Fibonacci number.""" if n == 0: return 0 elif n == 1: return 1 else: return fibonacci(n-1) + fibonacci(n-2) for n in range(n - 1, n + 1): return n - 1 + 1 return nfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "bayliner1980/Qwen3-0.6B-MoE-A0.5B"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto") messages = [{"role": "user", "content": "Write one sentence about the sea."}]ids = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_tensors="pt", return_dict=True).to(model.device)out = model.generate(**ids, max_new_tokens=128, do_sample=True, temperature=0.7, top_p=0.8)print(tok.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))