Hy-Memory-30B-A3B vs Qwen3.8-27B: The Specialist Is Not the Rival — It's the Replacement for the Job You're Already Paying a Generalist to Do

Nobody is choosing between Hy-Memory-30B-A3B and Qwen3.8-27B on capability, because they are not in the same race. Qwen3.8-27B is the lab's dense 27-billion-parameter vision-language model — Apache 2.0, 262,144-token native context, shipped on August 14, 2026, and served on OrcaRouter at $0.33 per million input tokens and $2.40 per million output. Hy-Memory-30B-A3B is the Hunyuan team's memory writer — 30B total with about 3B active, open-sourced October 15, 2026, and not served anywhere at all, including by us. One is a generalist you can rent by the token. The other is a component you have to host yourself. Benchmarks age and prices move; the failure record does not: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

The reason to put them side by side anyway is that Hy-Memory-30B-A3B exists to replace a specific thing, and that thing is very often a Qwen3.8-27B. Memory frameworks call a general-purpose model on every turn to decide what to keep. If you are running Qwen3.8-27B for that — and at $0.33/$2.40 it is a natural choice — then the honest question is not which model is better. It is whether a 3B-active specialist does the same job well enough to justify the hardware, and that question has a price tag attached that you can actually work out.

What Qwen3.8-27B is, in the terms a memory system cares about

Dense, which means every one of its parameters is read for every token. 64 layers at a hidden size of 5,120, with a 16-to-1 pattern alternating Gated DeltaNet linear-attention blocks and full-attention blocks — a design that is what makes a quarter-million-token window affordable to serve at all. It takes text, images and video in and returns text, which matters here for a mundane reason: a large share of what people want an agent to remember arrives as a screenshot or a document rather than as prose.

The vendor's own card claims 90.3 on LiveCodeBench v6, 89.2 on GPQA Diamond, 73.0 on Terminal-Bench 2.1, 84.3 on OSWorld-Verified and 79.5 on IFBench — all vendor-reported, unreproduced by us or anyone we can point to for the memory-writing workload specifically. The model has picked up independent placements since release, including a top-10 finish on a web-development arena leaderboard, but none of those measure whether it writes good memory. Nothing does yet.

What makes it the incumbent for this job is not any of those scores. It is that it follows instructions precisely, it accepts a JSON schema, it has a 262K window so you can hand it a long conversation without chunking gymnastics, and it is cheap enough to run on every turn.

The page above is the practical difference between these two models in one screen: Qwen3.8-27B has a rate and a route, and Hy-Memory-30B-A3B has neither.

What Hy-Memory-30B-A3B does differently

The Hy-Memory model is trained for exactly one task and its card says so without hedging: it replaces the general-purpose LLM inside Mem0, MemOS and Hy-Memory, deciding what information to retain, how precisely to represent it, and how new evidence should update a persistent bank. It is then evaluated on EverMemOS, a framework held out of training entirely — the one design decision in the release that most deserves credit, because cross-framework transfer is the entire premise.

Its training is two-stage, and the second stage is the interesting one. Memory On-Policy Distillation supplies dense supervision on memory states the student produced itself. Attributed Memory Reinforcement Learning then builds banks from rollouts, scores them chunk by chunk, penalises over-storage, and — this is the part that keeps training honest — carries the selected bank forward across windows rather than resetting between them. A deployed bank is never empty, so training it against an empty bank would be training it for a situation that does not exist.

Architecturally it is 48 layers, one dense block and 47 sparse MoE blocks, 128 experts at top-8 routing plus a shared expert, a 262,144-position context and 60.1 GB of BF16 weights across 51 shards. No quantised build was published.

Doing the arithmetic the cards don't do

Both models fire on every turn of a long conversation. That makes the write path a volume problem, and volume is where a specialist's smaller active-parameter count should show up.

Take a concrete, stated assumption rather than a measured one: a writer processes about 800 tokens of conversation per turn and emits about 150 tokens of structured memory. Over 1,000 turns that is 800,000 input tokens and 150,000 output tokens per user. On OrcaRouter's rate card for Qwen3.8-27B — the vendor's list price passed through at 0% markup, so $0.33 and $2.40 per million — that conversation costs roughly $0.26 in input and $0.36 in output, about $0.62, plus whatever the read path costs when an answer model opens the bank. Those are arithmetic on published rates, not a benchmark, and your tokens-per-turn will differ.

For Hy-Memory-30B-A3B the same 1,000 turns cost nothing per token and something considerably larger per month, because there is no rate card. The model is self-hosted, and self-hosting 60 GB of BF16 with no quantised alternative means an accelerator with the memory to hold it — before you know whether the 55.49 macro average holds up.

So the trade is not cheap against expensive. It is metered against capital, and the crossover point depends entirely on how many turns you actually write per month. A team running a few thousand conversations a month is almost certainly better off paying Qwen3.8-27B's per-token rate and never touching the hardware question. A team writing memory on millions of turns is running hardware either way, and the specialist becomes interesting.

The card above is where the arithmetic starts: there is a downloadable 27B checkpoint behind the $0.33 rate, and if your volume is high enough to justify running it yourself, that is a supported path with published quantisations behind it.

Three things the price comparison hides

Cost per token is the easiest axis and the least decisive one. Three others matter more.

Multimodality. Qwen3.8-27B takes images and video. Hy-Memory-30B-A3B's card specifies text input, English and Chinese. If your agent's memory includes screenshots, scanned documents or video frames, the generalist is doing work the specialist structurally cannot do, and no amount of cost advantage closes that.

Context. Both publish 262,144 positions, which is a genuine coincidence and a useful one — neither has an advantage in how much conversation you can hand it at once.

Published precision. Qwen3.8-27B ships as BF16 with an official FP8 build and GGUF conversions from several publishers, and OrcaRouter serves a block-FP8 version with the vision tower at full precision. Hy-Memory-30B-A3B ships BF16 and nothing else. That row decides whether you can evaluate it on a borrowed card or need to reserve one.

The routing answer, which is different depending on which model you pick

This is where the two paths separate cleanly, and it is worth being precise rather than promotional.

Qwen3.8-27B is in OrcaRouter's catalogue. You call it through the same API key as 200-odd other models, at the provider's list price with 0% markup — so a price cut from the provider is live here the same day rather than at the next billing cycle — with automatic failover if a backend degrades. For a memory harness, the useful part is the routing DSL: it composes several models into one call, which is how you run a writer and an answer model behind a single endpoint without building a second client. Model fusion, which runs a panel of models answering together, is the tool for working out empirically whether a specialist writer produces better banks than the generalist — run both against the same conversation, compare the answers, and stop arguing about it.

Hy-Memory-30B-A3B is not in our catalogue and we do not host it. The route to it is the checkpoint from the team and your own hardware. If you self-host it, a router still covers the rest of the loop — the hosted answer model that reads the bank back, the planning calls around it — and it lets you register your own endpoint alongside the hosted ones so failover and one-key access apply to both.

Who should do what

If you are running memory writing on a hosted generalist today, the honest first question is not whether to switch. It is how many turns you are actually writing per month. Below a threshold that only you can calculate, Qwen3.8-27B on a per-token rate is the rational choice and Hy-Memory-30B-A3B's hardware requirement is a distraction — particularly if your memory contains images, where the specialist has nothing to offer.

If you are writing memory at a volume where the per-token line has become a real budget item, Hy-Memory-30B-A3B is the first open model aimed squarely at that line, and the fact that it was tested against a held-out framework is the sign that someone thought about generalisation rather than about a demo. Wait for a quantised build. Wait for somebody outside the team to run the harness. Then decide.

And if you are simply curious whether a specialist beats a generalist on your own traffic, the experiment is cheap to set up and the answer will be specific to your data rather than to either card.

One last thing the card makes plain and the tables do not. Hy-Memory-30B-A3B is licensed Apache 2.0, which means the experiment above is one you are allowed to run commercially, on your own data, without a conversation with anybody's legal team. That is the strongest argument the model has, and it is independent of whether the 55.49 reproduces.