Juneo Lab's Laplace-1-Vector on a Mac: a 4B decision model in MLX
NetProjects
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-Vector is the 4B member of that family, built on Spark-X2.6-4B and released with open weights under the Apache-2.0 licence. We ported it to MLX so it runs on Macs with Apple Silicon. At 4B it is small enough to sit next to an agent on the same machine, so the data need not leave. We checked every build against Juneo Lab's original 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-Vector, 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
What the model does
You give Laplace-1-Vector 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 an 8.2 GB bf16 model: https://huggingface.co/JuneoAI/Laplace-1-Vector
Juneo Lab builds it around four decisions: allocating work, sharing information, acting and evaluating outcomes. Laplace-1-Vector is the smaller of the two Laplace-1 models; the larger is the 27B Laplace-1-Atlas, which we have also ported: https://netprojects.com/insights/juneo-laplace-1-atlas-decision-model-mlx-gguf. 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 on Hugging Face, converted from Juneo Lab's weights at revision 06cea7d. The quantised builds quantise the linear layers, affine with group 64, and keep the attention output gate in bf16.
| Hugging Face | Size |
|---|---|
| netprojects/Laplace-1-Vector-MLX-6bit (recommended) | 3.3 GB |
| netprojects/Laplace-1-Vector-MLX-8bit | 4.4 GB |
| netprojects/Laplace-1-Vector-MLX-4bit | 2.3 GB |
| netprojects/Laplace-1-Vector-MLX-bf16 | 8.2 GB |
We tested the builds with mlx 0.32.4 and mlx-lm 0.32.2. mlx-lm 0.32 and newer support this architecture, spark2_5, natively, so these builds do not need trust_remote_code. You need Python 3.11 or later; the macOS system Python is too old.
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 compared each build with Juneo Lab's own model code running in float32 on CPU, Transformers 4.57.1, on 11 prompts, including one of 3,313 tokens.
Measured on Laplace-1-Vector, 16 November 2026, MacBook Pro M5 Max (128 GB).
| Build | Same next token as the original | JevBench (231) | Same top answer as bf16 |
|---|---|---|---|
| MLX bf16 | 99.6% | 78.8% | Reference |
| MLX 8-bit | 98.9% | 79.7% | 229 / 231 |
| MLX 6-bit | 98.1% | 79.2% | 222 / 231 |
| MLX 4-bit | 90.4% | 78.4% | 210 / 231 |
The token-parity figures are teacher-forced over 753 positions. Cached decoding agrees with the full forward pass in every build. On 16 further candidate-scoring cases, every build picks the same candidate as the original.
JevBench is an open decision benchmark. Each of the 231 public items 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. Every build gets all 48 easy items right. On the 72 standard items, bf16 gets 68, 8-bit and 6-bit get 69, and 4-bit gets 67. On the 111 hard items, bf16 and 6-bit get 66, 8-bit gets 67, and 4-bit gets 66.
A total one or two items above bf16, as with the 8-bit and 6-bit builds, is noise, not an improvement. The 4-bit build's accuracy holds, but it changes the top answer on about 9% of JevBench items compared with bf16, which is why we recommend the 6-bit or 8-bit build for scoring and ranking work.
Juneo Lab reports 89.0% 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>=0.32"
mlx_lm.server --model netprojects/Laplace-1-Vector-MLX-6bitThe model card includes Juneo Lab's score_candidates helper with the same prompt, ported to mlx-lm. 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. The probabilities are relative to the candidates you supply, not a calibrated chance of being correct. For outcome prediction, send the environment rules as the system message and the state and action as JSON, with thinking off. The model card has the full commands: https://huggingface.co/netprojects/Laplace-1-Vector-MLX-6bit
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. Put the rules that govern the decision in the state, and include a hand-back or request-more-information option when the stakes call for one.
Honest limits
- All figures come from one machine.
- We ran only the public JevBench set, with our own prompt, plus 16 candidate-scoring cases and 11 prompts for token parity.
- We have no speed figures yet, and no memory figures beyond the file sizes above.
- We have not run Juneo Lab's other decision benchmarks, including AgentDecisionBench, COMPASS, OrgAccess, When2Call, JEVal and Coherence.
- The 4-bit build differs more from the original than the others.
Licence
Laplace-1-Vector, Spark-X2.6-4B 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 XHToken. If you use the model, Juneo Lab asks you to cite its Laplace-1 paper; the citation is on the model cards.
Links
- MLX 6-bit build (recommended): https://huggingface.co/netprojects/Laplace-1-Vector-MLX-6bit
- MLX 8-bit build: https://huggingface.co/netprojects/Laplace-1-Vector-MLX-8bit
- MLX 4-bit build: https://huggingface.co/netprojects/Laplace-1-Vector-MLX-4bit
- MLX bf16 build: https://huggingface.co/netprojects/Laplace-1-Vector-MLX-bf16
- NetProjects MLX builds collection: https://huggingface.co/collections/netprojects/netprojects-mlx-builds
- Juneo Lab's original: https://huggingface.co/JuneoAI/Laplace-1-Vector
- Juneo Lab's Laplace-1-Atlas on everyday computers: https://netprojects.com/insights/juneo-laplace-1-atlas-decision-model-mlx-gguf
- Also a decision model on a Mac, Fastino GLiDE: https://netprojects.com/insights/fastino-glide-decision-model-on-apple-silicon-mlx
- JEV-35B and AutoTrust-JEV-27B on a Mac: https://netprojects.com/insights/autotrust-jev-decision-models-on-apple-silicon-mlx
- All our AI ports: https://netprojects.com/ai
- How we use AI: https://netprojects.com/standards/ai
Always up to date. We keep these builds current on Hugging Face. When Juneo Lab releases a new version of Laplace-1-Vector, 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
Credits
Laplace-1-Vector is by Juneo Lab, built on Spark-X2.6-4B by XHToken, under the Apache-2.0 licence. The MLX builds are by NetProjects (Internet Projects Ltd). Not affiliated with Juneo Lab or XHToken.
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.