# Macaron-V1: four 1B adapters on a frozen 744B base

> Satyajit Ghana — Head of Engineering @ Inkers Technology
> canonical: https://ai.thesatyajit.com/articles/macaron-v1
> date: 2026-07-27
> tags: llm, lora, agents, open-weights, generative-ui, explainer
[Macaron-V1](https://huggingface.co/collections/mindlab-research/macaron-v1) is Mind Lab Research's agent
model family, released 2026-07-21 under MIT. The flagship, **Macaron-V1-Venti**, is described as a 748B model.
That number needs an asterisk immediately: **744B of it is a frozen GLM-5.2**, and Mind Lab's contribution is
**four 1B LoRA adapters** — about half a percent of the artifact.

That is the whole idea, and it is a genuinely different bet from how most post-training is done. Rather than
fine-tuning one monolithic model, take a strong open base, freeze it, and attach a small number of tiny
specialists. They call it **Mixture of LoRA (MoL)**.

## Mixture of LoRA

Four adapters, each 1B: `l0` Chat, `l1` Agent, `l2` Coding, `l3` GenUI. The routing detail is the neat part —
**`l0` is both the conversational backbone and the router**. It sees each new user request and dispatches it to
whichever specialist fits.

<MolRouter />

Note what routes and when. In a Mixture-of-Experts model, a router fires on *every token* at *every MoE layer*,
and the experts were built during pretraining — they are inseparable from the model. MoL routes **once per
request**, at the adapter level, and the thing being routed between is four swappable files sitting on a base
someone else trained. Ongoing reasoning and tool interaction stay inside the selected LoRA for the duration;
when a specialist finishes, its work is passed to the next as a concise summary rather than shared state.

The practical consequences are real. Specialization costs 1B parameters instead of a full fine-tune. Adapters
can be swapped, added, or updated independently. And when GLM-5.2 improves, you re-fit adapters rather than
retrain a 744B model. The cost is equally real: you inherit the base's ceiling, its licence obligations, and
its failure modes, and a request-level router cannot change its mind halfway through a turn the way per-token
routing implicitly can.

<Callout type="note">
**Update, 2026-08-03: the Tall size conflict looks like two different counts, not a contradiction.** Mind Lab's
blog calls Macaron-V1-Tall **50B**; the model card, the Hugging Face listing and Novita all say **36B**. The
Hugging Face API settles half of it — `safetensors.total` for `Macaron-V1-Tall` is exactly **35,951,822,704
(35.95B)**, which is the 36B figure and is the *base checkpoint on its own*. The blog's 50B is its own
decomposition of base plus adapters: 35B + 4 × 3.7B ≈ 49.8B. So the two numbers are measuring different things,
and Novita hedges the gap as "10~50B".

What I could not verify is the per-adapter figure. The published 35.95B total does not appear to include the
adapter weights, so I cannot confirm 3.7B each from the metadata — that number is Mind Lab's, not something I
measured. It is worth flagging because it implies Tall's adapters are roughly **3.7× the size of Venti's 1B
ones** on a base twenty times smaller: on Venti the specialization is about half a percent of the artifact, on
Tall closer to a tenth. If that holds, "Mixture of LoRA" means something quite different at the two scales, but
the evidence for it is currently a single line in a blog post.
</Callout>

## What the adapters actually buy

Most of the release table compares Macaron to Claude, GPT and Gemini. That is the least informative comparison
available, because it confounds the adapters with GLM-5.2's own strength. The controlled experiment is sitting
right there in the same table: **each variant against the frozen base it was built on**. That isolates the only
thing Mind Lab actually changed.

<AdapterLift />

The pattern is consistent and modest: roughly **+3 to +6 points** across chat, agent and coding work. Two rows
are inside noise — `T3-Bench` at +0.2 and SWE Atlas QnA at +0.6. And then UI4ABench jumps **+20.7** on Venti and
**+25.4** on Tall.

That outlier is the honest crux of the release. It is simultaneously the strongest evidence that a 1B adapter
can teach a frozen base a genuinely new skill, *and* the result most exposed to selection effects — UI4ABench is
Mind Lab's own benchmark, measuring generative UI, which is exactly the capability they built a dedicated
adapter for. Both things are true at once.

## Read the evaluation table twice

The published benchmark figure includes its own methodology notes, and they change how several rows should be
read.

<Figure
  src="/articles/macaron-v1/fig1.png"
  alt="Macaron V1 benchmark table across Chat, Agent, Coding and GenUI categories, comparing Macaron V1 Venti and Tall against GLM 5.2, GPT 5.5, Claude Opus 4.8, Gemini 3.1 Pro, Qwen 3.7 Max, Minimax M3 and Qwen3.6 35B-A3B, followed by per-benchmark evaluation protocol notes describing judge models, retry policies and scoring methods."
  caption="The full Macaron-V1 evaluation table — note the per-benchmark protocol notes underneath, which specify the judge model and retry policy for each row (Mind Lab Research, 2026)."
/>

Three things stand out:

- **The judges are other models, and one of them is the base.** ChatBench is scored by "a privately deployed
  GLM-5.2 judge" — and Venti *is* GLM-5.2 plus adapters. A model's own base evaluating its output is a conflict
  worth naming. Elsewhere the judge is a competitor: Claude Opus 4.6 on LivingBench, Claude Haiku 4.5 on
  PinchBench, GPT-5.4 on ClawGym, GLM-5.1 on VitaBench, Gemini 3.5 Flash scoring UI4ABench rubrics.
- **Several rows are best-of-N, not single-shot.** PinchBench reports "the best observed score". DeepSWE allows
  "up to three attempts, and report the best one". SWE Atlas QnA is pass@3. SWE-Bench Verified permits up to
  three retries on evaluation errors and reports the best successful attempt. Those are legitimate protocols,
  but they are not comparable to a single-trial number from another lab's report — and some competitor cells are
  marked as taken from leaderboards or the models' own reports.
- **Macaron does not lead everywhere.** Claude Opus 4.8 wins SWE Verified (88.6 vs 85.6) and SWE Atlas QnA (57.3
  vs 49.5). GPT-5.5 wins ClawGym (82.5 vs 77.7) and DeepSWE (70.0 vs 58.4). Qwen 3.7 Max wins VitaBench (61.2 vs
  60.0) and Gemini 3.1 Pro wins VitaBench2 (50.2 vs 46.0). Mind Lab says as much in its own post: "Coding is
  where we currently sit close to, rather than ahead of, the frontier."

The rows where Gemini, Qwen and Minimax collapse to 10.0–22.6 on DeepSWE and SWE Atlas QnA are almost certainly
harness incompatibility rather than capability — those evaluations run through Claude Code as the agent harness,
which is not neutral ground for every model.

## The infrastructure claims

Three systems are named, none with a technical report behind them yet:

- **MinT** — the post-training platform, claimed to support models up to a trillion parameters via adapter-only
  handoffs and a "million-scale adapter catalog". Adapter-only handoff is the load-bearing idea: if
  specialization is always a small file, you never move a 744B checkpoint between training stages.
- **MindForge** — an agentic RL framework built around discovery, expansion and update cycles against
  production harnesses.
- **LongStraw** — million-token RL, which works by evaluating a shared prompt once into a reusable resident
  state and then replaying only the response branches. For agentic RL where many rollouts share a long prefix,
  that is the obvious win, and it rhymes with the external KV-cache pooling in
  [Kimi K3](/articles/kimi-k3)'s RL infrastructure.

Macaron-V1 also ships a serving story: a
[Mixture-of-LoRA harness](https://github.com/MindLab-Research/Mixture-of-LoRA-Harness) that keeps an
OpenAI-compatible endpoint while adding the L0 router and same-request switching into the selected specialist,
plus [Macaron Artifacts](https://github.com/MindLab-Research/macaron-artifacts), a local WebUI and plugin that
runs inside Claude Code, Codex or Kimi Code.

## The take

The interesting claim in Macaron-V1 is architectural, not competitive: that request-level routing across a few
1B adapters on a frozen base is enough to build a specialized agent model, and that you can therefore treat a
frontier open-weight model as infrastructure rather than as something to fork. The base-versus-tuned comparison
supports a weaker version of that claim than the headline table does — a few points nearly everywhere, and one
large gain on the capability they purpose-built an adapter for.

What would settle it is the technical report, which the model card lists as "coming soon", along with the full
benchmark methodology. Until then this is a self-reported release with no third-party replication, several
best-of-N protocols, and its own base model sitting on the judging panel. The idea is worth watching. The
numbers are worth waiting on.

---

*Sources: the [Macaron-V1 collection](https://huggingface.co/collections/mindlab-research/macaron-v1) and the
[Macaron-V1-Venti](https://huggingface.co/mindlab-research/Macaron-V1-Venti) model card (architecture, adapter
roles, parameter counts, benchmark table), and Mind Lab's
[Introducing Macaron-V1](https://macaron.im/mindlab/research/introducing-macaron-v1) post (MinT, MindForge,
LongStraw, variant sizes). All benchmark numbers are Mind Lab's own, with per-benchmark judge models and
retry policies as annotated in their published table; no technical report has been released. The interactives
are mine.*