Marico's space

Qwen 3.8 27B Architecture (64 layers total)
══════════════════════════════════════════════════════ 16 MACRO-BLOCKS × { ┌─────────────────────────────────────────────┐ │ Gated DeltaNet → FFN (linear, 48V/16QK)│ ← Layer 1 │ Gated DeltaNet → FFN (linear, 48V/16QK)│ ← Layer 2 │ Gated DeltaNet → FFN (linear, 48V/16QK)│ ← Layer 3 │ Gated Attention → FFN (GQA, 24Q / 4KV) │ ← Layer 4 └─────────────────────────────────────────────┘ }
══════════════════════════════════════════════════════ + Native Vision Encoder (VLM) + Multi-Token Prediction heads (× D depths) + RoPE: 262,144 native context → 1M with scaling + Vocabulary: 248,320 (padded to 262,144) + Hidden dim: 5,120 | Head dim: 128 (DeltaNet) / 256 (Attn)
══════════════════════════════════════════════════════

S_t = (I - β_t · k_t · k_t^T) · S_{t-1} + β_t · v_t · k_t^T
o_t = S_t · q_t

L_total = L_next_token + (λ / D) × Σ(k=1 to D) L_MTP(k)

from openai import OpenAI
import os client = OpenAI( base_url="http://localhost:1234/v1", # LM Studio本地服务器
 api_key="lm-studio",
) def query_qwen38(prompt: str, effort: str = "low") -> dict: """ 用可配置的推理深度查询Qwen 3.8 27B。 Args: prompt: 用户消息 effort: "xhigh" | "medium" | "low" Returns: 带有"reasoning"和"answer"键的字典 """ completion = client.chat.completions.create( model="qwen3.8-27b", messages=[{"role": "user", "content": prompt}], reasoning_effort=effort, extra_body={ "chat_template_kwargs": { "enable_thinking": True, "preserve_thinking": True, # 保持推理在KV缓存中
 }, }, stream=True, ) reasoning_content = "" answer_content = "" is_answering = False for chunk in completion: delta = chunk.choices[0].delta # 捕获推理轨迹（思考Token）
 if hasattr(delta, "reasoning_content") and delta.reasoning_content: if not is_answering: print(delta.reasoning_content, end="", flush=True) reasoning_content += delta.reasoning_content # 捕获最终答案
 if hasattr(delta, "content") and delta.content: if not is_answering: print("\n" + "=" * 40 + "\n✅ 答案:\n" + "=" * 40) is_answering = True print(delta.content, end="", flush=True) answer_content += delta.content return {"reasoning": reasoning_content, "answer": answer_content} # 示例：简单编程任务用低努力度
result = query_qwen38( prompt="写一个Python函数来展平任意深度的嵌套列表。", effort="low"
) # 示例：难题用xhigh
result_hard = query_qwen38( prompt="用构造性论证证明存在无穷多个素数。", effort="xhigh"
)

def get_sampling_params(thinking: bool) -> dict: """ 返回Qwen 3.8 27B的最佳采样参数。 Qwen团队的官方建议。 """ if thinking: # 高温度用于探索多样化的推理链
 return { "temperature": 1.0, "top_p": 0.95, "top_k": 20, "presence_penalty": 0.0, # 推理时不惩罚重复Token
 } else: # 低温度用于专注、干净的生成
 return { "temperature": 0.7, "top_p": 0.80, "top_k": 20, "presence_penalty": 1.5, # 减少输出中的重复
 } def query_fast(messages: list, client: OpenAI) -> str: """非思考（最快）模式——适合RAG、分类、简单提取。""" params = get_sampling_params(thinking=False) top_k = params.pop("top_k") response = client.chat.completions.create( model="qwen3.8-27b", messages=messages, **params, extra_body={ "top_k": top_k, "chat_template_kwargs": {"enable_thinking": False}, }, ) return response.choices[0].message.content

这个任务对时间敏感（期望<5秒）？ └─ 是 → enable_thinking=False（非思考模式） └─ 否 → 这是多步骤Agentic任务吗？ └─ 是 → reasoning_effort="medium"起步 （xhigh可能导致分析瘫痪+重试循环） └─ 否 → 这是数学、形式推理或复杂代码生成吗？ └─ 是 → reasoning_effort="xhigh" └─ 否 → reasoning_effort="low"

# 默认：