hermes自用聚合api封装

一、分组(两组,用户自调成员)

  • 文组:纯文字推理,26 万上下文起。成员:v100 私有(custom/qwen3.8-flash-next-gyro-gyro-m)+ go 免费文字模型。Hermes 主模型 + 除图文外所有辅模型都指它
  • 图组:支持图片的模型 only,成员必须逐个实测验过(v100 大概率纯文字不进组)。Hermes 图片/语音识别辅模型指它
  • 百万组已取消:长上下文首字慢,不进日常轮询,读长文档任务单独调
  • 组内 strict-random 轮询(管并发/故障转移,不管单次提速);用户随时增删成员,不改 Hermes 侧

二、Hermes 接线

  • 主模型 → 文组;图片/语音辅模型 → 图组;其余辅模型 → 文组
  • 网关 combo 名即对外模型名,Hermes config 里三处分别指向两个组名即可

三、总架构图(EN boxes)

+=============================================================+
|              hermes + omniroute  (max2, ghost_net)        |
+=============================================================+

+------------------+    model=TEXT / VISION    +-------------------+
| hermes           |    Bearer gateway-key     | omniroute gateway |
| main      -> TEXT |                           | 3.8.53 single-ctr |
| aux img   -> VISION+-------------------------->| API :20128        |
| aux voice -> VISION|   http://omniroute:20128  | panel 127.0.0.1   |
| aux other -> TEXT  |   /v1                     | :20218            |
+------------------+                           +---------+---------+
                                                         | combos
                                               +---------+---------+
                                               | TEXT (strict-rand)|
                                               | custom/qwen3.8... |
                                               | opencode-go/*free |
                                               +---------+---------+
                                               | VISION(strict-rand|
                                               | (verified only)   |
                                               +------+------------+
                                                      |
                        +-----------------------------+------------------+
                        v                                                  v
+-------------------------------+         +----------------------------+
| v100 private compute          |         | opencode-go plan           |
| llama strata, Qwen3.8 125B    |         | native opencode-go type    |
| 262K ctx, text-only(?)        |         | https://opencode.ai/zen/go |
| http://v100:8080/v1           |         | /v1                        |
+-------------------------------+         +----------------------------+

RULES
1. round-robin != faster single reply; speed = right group
2. VISION members must pass image test one by one
3. long-ctx models stay out of daily rotation (slow first token)
4. user tunes membership; hermes side unchanged