How does Hy-Memory-30B-A3B score against Claude Opus 5?

How does Hy-Memory-30B-A3B score against Claude Opus 5?

Tencent's Hy-Memory-30B-A3B averages 55.49 across four memory harnesses (Mem0, Hy-Memory, MemOS and the held-out EverMemOS), ahead of Claude Opus 5 at 54.62 and DeepSeek V4.1 Flash at 53.64. The lead is small. It is nearly a tie on MemOS and the Hy-Memory harness, and clearest on EverMemOS. These are Tencent's own numbers, measured on fixed 20% subsamples of each benchmark.

What is Hy-Memory-30B-A3B?

It is a long-term-memory writer from Tencent Hunyuan with roughly 3B active parameters. A memory writer handles the write path in an external memory framework. It decides what to keep from a conversation, how precisely to phrase it, and how new evidence should update the stored memory bank.

The model is meant to replace the general-purpose LLM inside Mem0, MemOS and Hy-Memory, which covers Chat, Working and Coding scenarios. Training has two stages: Memory On-Policy Distillation, then Attributed Memory Reinforcement Learning. Both stages run inside the same deployed harness configurations. The weights, framework code and a technical report (arXiv 2610.22196) are public.

How was the benchmark set up?

The comparison changes one thing: the memory writer. Harness version, prompts, retrieval configuration, answer model and judge stay fixed. Each writer builds its own memory bank from an empty state, and decoding settings are matched where the APIs allow it.

Every benchmark runs on a fixed 20% subsample. Each harness score is the equal-weight mean of six metrics:

• LoCoMo Judge
• LongMemEval (LME)
• PersonaMem-v2 (PM-v2)
• Travel SPS
• Shopping PS
• Code Memory Bench (CMB), itself an average of two repository QA evaluations

The overall number is the equal-weight mean of the four harness scores. EverMemOS is held out from training, so it tests whether the writer transfers to an unseen framework.

What are the cross-harness results?

The ranking is the same as the overall order in most columns. Mem0 shows the widest spread among the top three: 54.02 for Hy-Memory-30B-A3B, 53.43 for Opus 5 and 49.73 for DeepSeek V4.1 Flash. Hy3 is notable on MemOS, where its 50.19 sits almost level with the top two.

Where does Hy-Memory-30B-A3B win, and where does it lose?

The per-benchmark tables show a more mixed picture than the averages.

Wins for the small model:

• Mem0 LongMemEval: 73.00 vs 66.00 for Opus 5.
• Mem0 Travel SPS: 54.51 vs 42.98 for Opus 5.
• Hy-Memory harness Travel SPS: 57.65, the best in that column.
• EverMemOS Code Memory Bench: 63.19 vs 49.53 for Opus 5 and 49.38 for DeepSeek V4.1 Flash.
• EverMemOS LoCoMo: 73.28, tied with Opus 5.

Wins for others:

• Mem0 LoCoMo: Opus 5 scores 69.47 against 59.54. That is the largest single-benchmark deficit in the tables.
• Mem0 Code Memory Bench: Opus 5 at 70.74 vs 67.13.
• Hy-Memory harness CMB: DeepSeek V4.1 Flash at 77.18 and Opus 5 at 75.70, vs 70.73.
• EverMemOS LongMemEval: Hy3 at 83.00 and DeepSeek V4.1 Flash at 82.83, vs 76.00.
• Shopping PS: Opus 5 and others lead on most harnesses, though every model scores low here (roughly 14 to 24 across all tables).

The pattern is that the small writer does well on the Travel and Code metrics and is less consistent on LoCoMo and LongMemEval. That is a reading of the tables, not a claim the report makes.

Does the held-out EverMemOS result show real transfer?

It is the most informative column. Training used Mem0, MemOS and Hy-Memory configurations. EverMemOS was never seen, and Hy-Memory-30B-A3B still scores 56.89, the highest in the table. Opus 5 reaches 54.04 and DeepSeek V4.1 Flash 55.47.

The margin over Opus 5 is 2.85 points, and most of it comes from CMB (63.19 vs 49.53). On LongMemEval within the same harness, the model trails DeepSeek V4.1 Flash by nearly seven points. So the transfer is real in aggregate but uneven by task.

The comparison with the base models is cleaner. On EverMemOS, Hy3-A3B Base scores 50.98 and Qwen3-30B-A3B scores 46.57, so the trained writer gains about six and ten points respectively. The model card does not include an ablation separating the OPD stage from the RL stage in these tables, so the contribution of each stage is not visible here.

How much should you trust these numbers?

Several caveats apply.

• Vendor-reported. Tencent ran the evaluation on its own model and its own Hy-Memory harness. Independent replication is not in the sources.
• 20% subsamples. Each benchmark is a fixed fraction of the full set. Small test sets make single-benchmark differences noisy, and values like 73.00 and 89.00 on LME suggest coarse granularity.
• Tiny margins. A 0.02 gap on Hy-Memory and 0.01 on MemOS is a tie. The overall lead over Opus 5 is under one point.
• Equal weighting. Averaging six metrics hides large swings, like the Mem0 LoCoMo gap. If your product depends on conversational recall, that column matters more than the mean.
• Fixed downstream components. The answer model, judge and retrieval do not change, so the scores measure the writer's effect only inside these configurations.

A fair summary is that a ~3B-active open-weight model matches frontier-class writers on this benchmark matrix. That is the claim the data supports. It does not show that the model is better in general.

How do you run Hy-Memory-30B-A3B?

The model card recommends serving with vLLM (tested with v0.31.0) using the served model name hy-memory, then calling the OpenAI-compatible API. Its example sends a system prompt asking for explicit user facts only, with a JSON object containing a memories list. It sets temperature to 0.0, a 512-token cap, and no-think mode:

extra_body={
    "chat_template_kwargs": {
        "reasoning_effort": "no_think",
    },
}

For production use, the intended path is dropping it into Mem0, MemOS or Hy-Memory in place of the general LLM. The repository includes framework code and evaluation utilities.