Juneo Lab's Laplace-1-Atlas on everyday computers: a 27B decision model in MLX and GGUF

NetProjects12 Nov 2026

Many decisions inside an organisation come down to a short list of options. Who should own this task? May this person see this record? Should an agent run this action, or ask first? Does this piece of work meet the standard? A model that only has to choose can give a probability for every option instead of writing an answer, and those probabilities tell you how sure it is.

Juneo Lab's Laplace-1-Atlas is built for exactly that. It is a 27B decision model on Qwen3.8-27B, released with open weights under the Apache-2.0 licence. We ported it so it runs on everyday computers: MLX builds for Macs with Apple Silicon, and GGUF builds for llama.cpp and the tools built on it, such as Ollama and LM Studio, on Windows, Linux and Mac. We checked every build against the full-precision weights and published them openly on Hugging Face.

Always up to date. We keep these builds current on Hugging Face. When Juneo Lab releases a new version of Laplace-1-Atlas, we rebuild, re-test and publish it there, so the links on this page always point to the latest version. The numbers in this article were measured on the version named in each section; the model cards on Hugging Face carry the current figures.

Hugging Face: https://huggingface.co/netprojects
NetProjects MLX builds collection: https://huggingface.co/collections/netprojects/netprojects-mlx-builds
NetProjects GGUF builds collection: https://huggingface.co/collections/netprojects/netprojects-gguf-builds

What the model does

You give Laplace-1-Atlas a state, a question and a list of candidates. It reads the candidates as letters, and the probabilities of those letters at the next token give a distribution over exactly the candidates you supplied, in one forward pass, with no sampling and no parsing. It can also predict the next state of an environment you describe, as text. Juneo Lab's original is a 54.7 GB bf16 model: https://huggingface.co/JuneoAI/Laplace-1-Atlas

Juneo Lab builds it around four decisions: allocating work, sharing information, acting and evaluating outcomes. Laplace-1-Atlas is the larger of the two Laplace-1 models; the other is the 4B Laplace-1-Vector. It sits next to the other decision models we have ported: Fastino's GLiDE 27B, AutoTrust AI's JEV models and H2O-Lightning-4B.

What we released

Four MLX builds and three GGUF files on Hugging Face. The MLX builds quantise the language model and keep the vision tower in bf16. The GGUF repository also holds a 0.9 GB vision projector, needed only for images.

Hugging Face	Size
netprojects/Laplace-1-Atlas-MLX-6bit (recommended for Macs)	22.8 GB
netprojects/Laplace-1-Atlas-MLX-8bit	29.5 GB
netprojects/Laplace-1-Atlas-MLX-4bit	16.1 GB
netprojects/Laplace-1-Atlas-MLX-bf16	54.7 GB
netprojects/Laplace-1-Atlas-GGUF, Q6_K (recommended for llama.cpp)	22.1 GB
netprojects/Laplace-1-Atlas-GGUF, Q8_0	28.6 GB
netprojects/Laplace-1-Atlas-GGUF, Q4_K_M	16.5 GB

We tested the MLX builds with mlx 0.32.4, mlx-lm 0.32.2 and mlx-vlm 0.7.11, and the GGUF files with llama.cpp build 11960.

How well it matches the original

We first checked that the weight files we downloaded from Juneo Lab match the hashes published on Hugging Face. Then we ran every build on the 231 public items of JevBench, an open decision benchmark. Each item is one candidate-scoring pass with thinking off, and the highest-probability candidate is compared with the expected answer. The bf16 MLX build, the full-precision weights, is the reference for the others.

Measured on Laplace-1-Atlas, 11 November 2026, MacBook Pro M5 Max (128 GB).

Build	JevBench (231)	Same top answer as bf16	KL, mean	Peak memory
MLX bf16	91.8%	Reference	Reference	53.2 GB
MLX 8-bit	92.2%	230 / 231	0.006	29.2 GB
MLX 6-bit	91.8%	231 / 231	0.008	23.0 GB
MLX 4-bit	91.3%	224 / 231	0.079	16.7 GB
GGUF Q8_0	92.2%	230 / 231	0.003	Not measured
GGUF Q6_K	91.8%	231 / 231	0.009	Not measured
GGUF Q4_K_M	92.6%	227 / 231	0.020	Not measured

KL measures how far a build's probabilities drift from bf16; lower is closer. Every build gets all 48 easy items and 71 of the 72 standard items right; the differences are all in the 111 hard items, where bf16 gets 93. The 6-bit MLX build and the Q6_K GGUF file give the same top answer as bf16 on every item at less than half the size, which is why we recommend them. A total one or two items above bf16, as with the 8-bit and Q4_K_M builds, is noise, not an improvement. The 4-bit builds differ the most from bf16.

On 24 further decision cases we wrote for this check, every MLX build picks the same candidate as bf16. The 4-bit, 6-bit and 8-bit MLX builds each describe a test image correctly in mlx-vlm.

Juneo Lab reports 93.2% on JevBench for the original. Our runs use our own prompt and settings, so our figures are not a reproduction of Juneo Lab's.

Try it

On a Mac with Apple Silicon, with mlx-lm:

pip install -U mlx-lm
mlx_lm.server --model netprojects/Laplace-1-Atlas-MLX-6bit

On Windows, Linux or a Mac, with llama.cpp:

llama-server -hf netprojects/Laplace-1-Atlas-GGUF:Q6_K --port 8080 -c 8192

Both model cards include Juneo Lab's score_candidates helper with the same prompt, ported to mlx-lm and to llama-server. A request looks like this, from the model card example:

score_candidates(
    state={"permissions": ["read_records"], "request": "Delete a record."},
    question="Which decision is supported by the supplied permissions?",
    choices={"allow": "Allow the request.", "deny": "Deny the request."},
)

It returns the choice and a probability for each candidate. For outcome prediction, send the environment rules as the system message and the state and action as JSON, with thinking off. The model cards have the full commands:

Laplace advises; your system decides

Juneo Lab is clear that the model scores and never executes. Keep deterministic permission checks, authentication and audit logging in place, calibrate thresholds on your own decisions before automating on them, and keep a person on the final decision for irreversible, safety-critical, financial, medical or legal actions. The probabilities are relative to the candidates you supply, not a calibrated chance of being correct, so put the rules that govern the decision in the state.

Honest limits

All figures come from one machine.
We ran only the public JevBench set, with our own prompt, plus 24 decision cases of our own.
We have no speed figures yet, and no memory figures for the GGUF files.
We tested the GGUF files with llama.cpp's own server, not in Ollama or LM Studio.
Juneo Lab reports its decision results on text. We checked images only by asking each quantised MLX build to describe one test image.
The 4-bit builds differ more from the original than the others.

Licence

Laplace-1-Atlas, Qwen3.8-27B and our builds are all under the Apache-2.0 licence. Each of our repositories carries the upstream licence and a notice that describes the conversion and credits Juneo Lab and Alibaba Cloud. If you use the model, Juneo Lab asks you to cite its Laplace-1 paper; the citation is on the model cards.

Credits

Laplace-1-Atlas is by Juneo Lab, built on Qwen3.8-27B by Alibaba Cloud's Qwen team, under the Apache-2.0 licence. The MLX and GGUF builds are by NetProjects (Internet Projects Ltd). Not affiliated with Juneo Lab or Alibaba Cloud.

This article was drafted with AI assistance, then edited and approved by people on our team. You can read more about how we use AI at https://netprojects.com/standards/ai.