Sign in

Head to head · Open-weight models you run yourself · October 2026 research run

Hy-Memory-30B-A3B vs Hy3

Hy3 scores 62.6 (B) on agent readiness against Hy-Memory-30B-A3B's 61 (C), and leads in 1 of 4 scored categories. Hy-Memory-30B-A3B leads on reach and documentation and upkeep. Both do open-weight models you run yourself.

Best open-weight models you download and run yourself for AI agents · All 4969 open weights comparisons

Hy-Memory-30B-A3BTencent (Hy team) · Open-weight model C61/100#62 of 135 open-weight models Hy3Tencent (Hy team) · Open-weight model B62.6/100#28 of 135 open-weight models

Which one, for what

Hy-Memory-30B-A3B C

Best for Agents with an external memory framework that want a local model for the write path on a 24 GB card

Ahead on

  • Reach, 85 against 70
  • Documentation and upkeep, 68 against 56

Watch for

Not a general assistant or a tool caller

Hy3 B

Best for Teams with an 8-GPU server who want an agent-tuned MoE model under plain Apache 2.0

Ahead on

  • Agent interface, 75 against 65

Watch for

597.6 GB in BF16, 299.9 GB in FP8 and 168.5 GB at the smallest 4-bit build

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Reach Hy-Memory-30B-A3B +15

Hy-Memory-30B-A3B 85

Memory 25, Tencent's own Q4_K_M GGUF is 18.24 GB, and 18.24 GB × 1.2 = 21.9 GB, which fits 24 GB but not 16 GB. Its Q3_K_M file and the two files with experts at 1 or 2 bits are below 4-bit precision and don't count. Runtimes 20, hy_v3 model code in released vLLM (HYV3ForCausalLM in the v0.33.0 registry), SGLang v0.5.23 (hunyuan_v3.py), Transformers v5.22.2 and llama.cpp b12242, 5 each. Builds 15, Tencent publishes an FP8 repository and a GGUF repository with Q4_K_M, Q6_K and Q8_0 files in its own organisation, so quantised builds at two precisions or more. Terms 15, Apache 2.0, with only a preamble naming Tencent as copyright holder above the licence text. Download 10, gated is false for the main, FP8 and GGUF repositories on 10 November 2026, so the files download with no account or terms to accept. Hosted 0, Hugging Face lists no inference provider for the model (an empty provider mapping on 10 November 2026), and the card names no API. 25 + 20 + 15 + 15 + 10 + 0 = 85.

Hy3 70

Memory earns 5, because the smallest build of 4-bit precision or better is AngelSlim's GPTQ-Int4 of 168.5 GB, about 202 GB with the fifth added, which is within 256 GB but beyond any single workstation. The 1-bit IQ1_M GGUF fits our reference machine but sits below the 4-bit line. Runtimes earn 20, for model code in released Transformers, vLLM, SGLang and llama.cpp; Ollama and LM Studio have no entry for it. Builds earn the full 15, with Tencent's FP8 repository and AngelSlim's GPTQ and GGUF builds. Terms earn 15 for plain Apache 2.0. Download earns 10, with no gate and no account. Hosted earns 5, since Featherless AI and DeepInfra serve it live through Hugging Face.

Documentation and upkeep Hy-Memory-30B-A3B +12

Hy-Memory-30B-A3B 68

Model card 8, Intended use (the write path of Mem0, MemOS and Hy-Memory), parameters (30B, about 3B active) and languages (English and Chinese, in the metadata) are there. Training data 5, A general description of the two training stages inside the harnesses, with no datasets, sizes or cut-off. Last change 20, the last commit, on 16 October 2026, added the licence and the card, 25 days before this check. Responsiveness 20, Of the 4 discussions that aren't pull requests (16 to 30 October 2026), 3 have a reply from a tencent organisation member. Version policy 5, A first release, with nothing replaced. File health 10, Config, tokenizer and template agree on the architecture, the context and the thinking levels. 8 + 5 + 20 + 20 + 5 + 10 = 68.

Hy3 56

The model card earns 11, covering intended use, architecture, parameters and context but not limitations, languages or a knowledge cut-off. Training data earns 5 for a general statement with no cut-off. The last change earns 10, a config.json edit on 20 July and AngelSlim uploads on 21 July, within 180 days on 2 November 2026; the line drops to 5 after 17 January 2027 unless the files change. Responsiveness earns 10, with Tencent replies on 7 of 19 discussions and none since 22 July. Version policy earns the full 10, since the preview stays in its own repository and the card dates it and says what changed. File health earns 10, with files that agree and no broken-file report waiting.

Agent interface Hy3 +10

Hy-Memory-30B-A3B 65

Chat template 20, chat_template.jinja in the repository. Tool-calling format 10, the card doesn't mention tools. Trained tool calling 10, the template can render tool calls, but the card presents the model as a memory writer that returns memories as text or JSON and says nothing of calling tools. Structured output 5, the card's quickstart asks for a JSON object with a memories list in the system prompt and shows nothing else, no response_format and no schema, so JSON output is shown by example without a documented way to constrain it. Reasoning control 10, the card passes reasoning_effort as no_think per request through chat_template_kwargs, and the template takes high, low or no_think. Context 10, 262,144 tokens native (max_position_embeddings in config.json). 20 + 10 + 10 + 5 + 10 + 10 = 65.

Hy3 75

The chat template earns 20, a chat_template.jinja in the repository. The tool-calling format earns 10, because the card names the parsers but leaves the format itself to the template and SGLang's cookbook. Trained tool calling earns 20, as the template takes several calls in one block. Structured output earns 5 for the card's claim of stable output formats under constraints, with no word on how to ask for JSON. Reasoning control earns 10, with effort set per request and switchable off, and reasoning returned separately through the named parsers, as the linked vLLM recipe shows. Context earns 10 for 262,144 tokens native.

Score by category

CategoryWeight this runHy-Memory-30B-A3BHy3Edge
Capability35%pendingpendingpendingnot scored in this run
Reach25%258570Hy-Memory-30B-A3B +15
Risk20%206060even
Agent interface10%106575Hy3 +10
Documentation and upkeep10%106856Hy-Memory-30B-A3B +12
Negative events≤1500
TotalC 61/100B 62.6/100

Facts side by side

FactHy-Memory-30B-A3BHy3
KindOpen-weight modelOpen-weight model
VendorTencent (Hy team)Tencent (Hy team)
Hosted endpointno (local only)no (local only)
Transports
AuthNoneNone
PricingFreeFree
x402nono
LicenceApache-2.0Apache-2.0
Read-only variant documentednono
llms.txtnono
Last release2026-10-162026-07-20
Terms last updatedno document linkedno document linked
Privacy policy last updatedno document linkedno document linked
Agent reviewsnone2.8/5 (4)

Verdicts

Hy-Memory-30B-A3B

A specialist that writes an agent's long-term memory, built to slot into Mem0, MemOS or Tencent's own framework in place of a general model. It fits a 24 GB card from Tencent's own GGUF, under plain Apache 2.0, and Tencent answers on the repository. Its results are Tencent's own and nobody has rerun them.

Hy3

Hy3 is a 295B mixture-of-experts model with 21B active, trained for several tool calls a turn, with per-request reasoning effort and 256K of context. It needs an 8-GPU server for FP8 or about 200 GB for a 4-bit build. The main risks are no safety evaluation or agent guidance, and no file changes or replies since July.

Before you call either

Hy-Memory-30B-A3B

  1. Use it as the writer model in Mem0, MemOS or Hy-Memory, not as the model that answers the user
  2. Start it with llama-server -hf tencent/Hy-Memory-30B-A3B-GGUF:Q4_K_M --jinja, or with vLLM under the served name hy-memory as the card does
  3. Pass reasoning_effort as no_think in chat_template_kwargs and temperature 0, as the card's example does
  4. Ask for a JSON object in the system prompt and validate it yourself, since no schema constraint is documented
  5. Treat what it stores as unverified. It keeps what a conversation says, true or not

Hy3

  1. Start it with vLLM on 8 GPUs, vllm serve tencent/Hy3-FP8 --tensor-parallel-size 8 --tool-call-parser hy_v3 --reasoning-parser hy_v3 --enable-auto-tool-choice --served-model-name hy3
  2. In SGLang use --tool-call-parser hunyuan --reasoning-parser hunyuan, as the card shows
  3. Pass reasoning_effort (no_think, low or high) in chat_template_kwargs; the default is no_think, and any other value raises a template error
  4. With the vLLM reasoning parser the chain of thought comes back in message.reasoning and the answer in message.content
  5. Earlier thinking stays in the prompt when tools are present; the template keys this on preserved_thinking, not the interleaved_thinking named in vLLM's recipe
  6. Use temperature 0.9 and top_p 1.0, the card's recommended sampling
  7. On a 128 GB machine only AngelSlim's IQ1_M GGUF fits, llama-server -m Hy3-IQ1_M.gguf -ctk q8_0 -ctv q8_0 -fa on -c 65536, with llama.cpp b11768 or later; at 89.4 GB × 1.2 it sits right at the 107.4 GB ceiling, so shorten the context if it fails to load, and expect a 1-bit build to fall short of Tencent's numbers
  8. Without the hardware, call it through Featherless AI or DeepInfra via Hugging Face's inference providers

Questions

Which is better for AI agents, Hy-Memory-30B-A3B or Hy3?

Hy3 scores 62.6 (B) on agent readiness against Hy-Memory-30B-A3B's 61 (C), and leads in 1 of 4 scored categories. Hy-Memory-30B-A3B leads on reach and documentation and upkeep.

Other comparisons with Hy-Memory-30B-A3B or Hy3

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.

Get an auditSee a sample report