Model release ·
x402 Multi-Endpoint Selection Head
A compact decision model that routes constrained, multi-step x402 service flows.
Download weightsThe name
x402
- M
Multi-
- E
Endpoint
- S
Selection
- H
Head
- Complete-flow exact match
- 80.0%
- Complete-flow exact match
- Active-service accuracy
- 97.5%
- Active-service accuracy
- Post-policy violations
- 0
- Post-policy violations
- Per full x402 flow, NVIDIA GPU
- 0.31 s
- Per full x402 flow, NVIDIA GPU
Overview
Models can call tools. xMESH decides the flow.
A useful x402 request rarely ends with one endpoint. It may need to acquire material, extract it, enrich it, analyze it, and deliver a result, all inside a budget and an approval policy.
xMESH scores typed choices instead of generating prose. It chooses only from live candidates returned by search, while deterministic code controls wallets, prices, schemas, and execution.
One request · four stages
- Acquire
- Extract
- Analyze
- Deliver
Extract → pdf.text
- 01
Retrieve
Fetch current endpoints, prices, schemas, and availability.
- 02xMESH
Decide
Score stage gates and choose only from supplied candidates.
- 03
Validate
Enforce budget, network, approval, and feasibility policy.
- 04
Execute
Let trusted code pay, call, record the outcome, and continue.
Steps 01, 03, and 04 are deterministic code. Only step 02 is xMESH.
A field of priced endpoints. One right choice per stage.
Model details
x402 Multi-Endpoint Selection Head
Singularity xMESH 0.8B
singularity-layer/Singularity-xMESH-0.8B
- Full name
- x402 Multi-Endpoint Selection Head (xMESH)
- Release
- Singularity xMESH 0.8B
- Model ID
- singularity-layer/Singularity-xMESH-0.8B
- Type
- Typed decision model with a calibrated Clef-style joint decision head
- Base model
- Qwen/Qwen3.5-0.8B, rank-16 adaptation
- Parameters
- 873M (BF16)
- Context
- 2,048 tokens validated
- Weights
- Safetensors 1.83 GB · BF16 GGUF 1.57 GB · Q8 GGUF 802 MB
- Runtime
- llama.cpp b11719 · CPU, Apple silicon GPU, NVIDIA GPU (Vulkan)
- Training data
- AgentToolDecisions-180K (mixed CC BY 4.0, Apache-2.0, MIT) plus Singularity-curated x402 rows
- License
- Apache-2.0
- Released
- October 18, 2026
Get the weights
View on Hugging FaceThree builds. Same decisions.
Universal · Singularity Node · CPU and GPU
$ huggingface-cli download singularity-layer/Singularity-xMESH-0.8B \
Singularity-xMESH-0.8B-BF16.gguf --local-dir .sha256 9f4c08f19c4d203059bc74ed2754e800ec472d7276d8ddb2321480e86d392fff
80 frozen x402 flows · every decision right
0.0%
Each square is one real request. xMESH got every gate, service, and global decision right in 64 of them.
- Jev
- 11.25% · 9/80
- Laya
- 0% · 0/80
Evaluation
Complete x402 flows, judged in one pass.
Every model received the same 80 frozen x402 payloads. Complete-flow accuracy counts a case only when every gate, service selection, and global decision in it is correct.
Frozen x402 FlowBench · 80 payloads
Accuracy by decision layer
Complete flow
Every stage, service, and global decision is correct.
- xMESH 0.8B
- 80%
- Jev
- 11.25%
- Laya
- 0%
Stage gates
Whether each step is required for the requested outcome.
- xMESH 0.8B
- 96.25%
- Jev
- 86.5%
- Laya
- 78.5%
Active service
The selected service is correct when that stage should run.
- xMESH 0.8B
- 97.5%
- Jev
- 100%
- Laya
- 85.63%
Global decision
Clarification and end-to-end feasibility decisions are correct.
- xMESH 0.8B
- 96.25%
- Jev
- 37.5%
- Laya
- 35.63%
Jev ran through its hosted API; xMESH and Laya ran locally. Accuracy is comparable across identical payloads. Latency is not hardware-equivalent.
Single-tool selection · 300 unseen cases
One request, four candidate tools, one right answer
Release weights, model only, before the policy layer. On par with Jev, far ahead of Laya.
Footprint · Apple M4
2.2GB
Peak memory while deciding
- Active memory
- 1.45 GB
- BF16 weights
- 1.57 GB
Reliability
Steady answers, fast.
Shuffle the options and xMESH gives the same answer, 600 times out of 600. It handles choice, yes-or-no, and score questions in one pass, and ten questions cost about 85 ms each on an Apple M4.
Even on a general decision benchmark it never trained on, it scores 11.6 points above base Laya.
Option order · seeded shuffle
0 / 600
Shuffle the options and ask again. Not one of 600 choice answers changed.
Three question types, one pass
Five-question case · M4 · Q8
1.20s
Median for a full typed-decisions case
Question scaling · Apple M4
One forward pass, more decisions
Each slice is one question answered in the same pass. Ten questions cost about half as much each as one.
−49%cost per question, 1 → 10
1 question
168 ms each
5 questions
89.7 ms each
10 questions
84.7 ms each
Questions per pass · ms per question
Fixed short state, three options per question, median of 20 sequential runs on an Apple M4.
Measured locally with llama.cpp b11719 on the immutable Q8 artifact, LocalLLaMA/typed-decisions test split. Laya figures are the ones published by the Laya benchmark page.
Training
Trained on Singularity Training Lab.
A Qwen3.5 0.8B backbone with a calibrated decision head, trained end to end in Training Lab. Frozen evaluation cases never touched training.
Training runs v6 → v11
From v6 to the released v11
- Service accuracy
- 82% → 94%
- Median latency · dev runs
- 2.90s → 1.39s
Development holdout, 40 flows. The release score is the frozen evaluation above.
Hardware
Runs everywhere. Decides the same.
Runs on CPU everywhere, and on Apple silicon and NVIDIA GPUs with the same decisions. On an NVIDIA RTX A4000 a full x402 flow takes 0.31 seconds. Singularity Node picks the GPU path automatically.
Decision parity · 960 frozen decisions
Same decisions on every chip.
llama.cpp b11719
CPU
macOSLinuxWindows
960/960 same as CPU · Reference
3.1 s
Apple M4
Apple silicon GPU
Apple M4 · Metal
959/960 same as CPU · One near-tie, about 50.1% vs 49.9%
0.82 s
Measured under load
NVIDIA GPU
RTX A4000 · Vulkan build
960/960 same as CPU · Every decision identical
0.31 s
Median per flow
Live on the Grid
Already serving on the Singularity Grid.
- 01Singularity Desktop v1.10.7 runs it locally on your GPU.
- 02One click serves it on the Grid. A Mac was the first node live.
Available now
Download xMESH.
Build the flow.
singularity-layer/Singularity-xMESH-0.8B
Methodology
Every number comes with a receipt.
How the evaluation was run, what each score means, and the hashes to check the exact weights yourself.
Evaluation receipt
Singularity xMESH 0.8B · October 18, 2026
Weights frozen first01
The standalone export was fixed before one scored run on the frozen 80-case suite. Nothing changed after.
Complete flow means all of it02
A case counts only when every gate, service, clarification, and feasibility decision is right.
Same payloads, not same hardware03
Accuracy compares like for like. Latency is deployment context only.
Public benchmark, pinned04
LocalLLaMA/typed-decisions revision 7189f46b, full all/test split, no policy wrapper.
Order cannot sway it05
One seeded shuffle per choice question; option IDs moved with their descriptions.
Every file checksummed06
MANIFEST.json covers all 47 published files; the model's own verifier passes on a public download.
- BF16 GGUF sha256
- 9f4c08f19c4d…0e86d392fff
- Q8 GGUF sha256
- 41b36270b67f…96e2729fd75
- Runtime
- llama.cpp b11719
- Post-policy violations
- 0