🧩 Model Card: hy-memory:30b-a3b

• Type: Text-to-Text (long-term memory writer, MoE, ~3B active parameters)
• Think: No (the template’s reasoning_effort defaults to no_think)
• Tool Calling Support: No
• Base Model: tencent/Hy-Memory-30B-A3B (Apache-2.0)
• Quantization: Q4_K (experts and projections), Q8_0 (lm_head)
• Max Context Length: 16k tokens
• Default Context Length: 16k tokens
• Host memory: about 19 GB for the weights (pinned in NPU-visible memory, not reclaimable) plus 1–2 GB of working memory, and up to about 4 GB more for the prefix cache; 32 GB or more of system RAM is recommended.

▶️ Run with FastFlowLM in PowerShell:

flm run hy-memory:30b-a3b

📖 Usage

Hy-Memory is the language model behind external memory frameworks (Mem0, MemOS, Hy-Memory). It receives one extract / reconcile prompt at a time and answers with a JSON object or array, so FLM serves it with greedy decoding by default. Every request is a whole conversation; the prefix cache reuses the part that matches an earlier request.

flm serve hy-memory:30b-a3b

{
  "model": "hy-memory:30b-a3b",
  "messages": [
    {"role": "user", "content": "Extract facts as a JSON array: Alice moved to Paris in 2021 and works at the Louvre."}
  ]
}

The chat template reads reasoning_effort (no_think, low or high; default no_think). Images and audio are ignored: the model is text only.

⚠️ The prompt is prefilled at about 200 tokens/s, so a full 16k-token prompt takes over a minute to its first token. Keep memory-writer prompts short, or reuse a long shared system prompt through the prefix cache.