GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
singularity-layer
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
singularity-layer
Model Tree
Qwen/Qwen3.5-0.8B
Input Modalities
Output Modalities
Supported Functionality
Input: an intent, hard constraints, and a small candidate set retrieved from a current x402 catalog. The validated runtime contract is 2,048 tokens.
Output: candidate selection or abstention, calibrated confidence, approval requirement, network compatibility, and cost level. Deterministic policy validates stage gates and flow feasibility before any paid request. Service choices remain model decisions.
The general decision warm-up used licensed rows from MaziyarPanahi/AgentToolDecisions-180K. That dataset retains mixed CC BY 4.0, Apache-2.0, and MIT source terms per row; see THIRD_PARTY_NOTICES.md. Singularity-curated x402 rows supplied the routing specialization. Frozen evaluation cases and resources were excluded from training.
| Hardware | Status |
|---|---|
| Apple silicon MPS | Verified on an M4 with 16 GB unified memory |
| CPU | Verified; portable but slower |
| llama.cpp Metal | Verified with the Q8 GGUF on an Apple M4 |
| llama.cpp Windows/Linux x64 CPU | Verified; native canaries passed 9/9 expected decisions on both platforms |
| NVIDIA CUDA | PyTorch runtime support only; Singularity Node GPU offload is disabled pending decision-parity canaries |
| AMD ROCm | Not verified |
The checkpoint contains 873M BF16 parameters. An 8 GB system is the practical minimum; 16 GB is recommended for comfortable local loading and inference.
Singularity-xMESH-0.8B-BF16.gguf is the universal Singularity Node artifact. It is 1,571,529,600 bytes with SHA-256 9f4c08f19c4d203059bc74ed2754e800ec472d7276d8ddb2321480e86d392fff and requires llama.cpp b11719 or newer for native Clef /v1/systemone support.
Singularity-xMESH-0.8B-Q8_0.gguf is an optional 841,069,440-byte accelerated artifact with SHA-256 41b36270b67f6c595e98cf4f29d6a0d13f0c19daa37b75741cd8296e2729fd75. It preserved all 960 decisions across the frozen 80-case Metal export-parity run, but llama.cpp b11719 crashes while initializing this quantized Clef architecture on CPU. Singularity Node therefore uses BF16 for its cross-platform contract. Q4_K_M was rejected because it changed 16 decisions and preserved only 65 of 80 complete flows.
The standalone export was fixed before one evaluation on the existing frozen 80-case x402 suite. No weights, prompts, or policy rules were changed after seeing these results.
| Evaluation, 80 cases | xMESH standalone | Jev | Laya |
|---|---|---|---|
| Complete-flow exact match | 80.0% | 11.25% | 0.0% |
| Stage-gate accuracy | 96.25% | 86.5% | 78.5% |
| Active-service accuracy | 97.5% | 100.0% | 85.63% |
| Global-decision accuracy | 96.25% | 37.5% |
Jev is a hosted API and Laya used a different local runtime, so latency is not a hardware-equivalent comparison. The standalone CPU validation measured 6.45 seconds median per full multi-field flow on an Apple M4. Optimized serving runtimes can be substantially faster.
This is a boundary test, not the model's specialization. xMESH was evaluated zero-shot on LocalLLaMA/typed-decisions revision 7189f46b02248bd2569df7325c9df8e3b24e90a7, complete all/test split: 400 cases and 2,000 decisions.
| Model | Condition | Accuracy | Soft accuracy | Brier ↓ | ECE ↓ | Score MAE ↓ |
|---|---|---|---|---|---|---|
| xMESH 0.8B | zero-shot Q8 | 0.478 | 0.415 | 0.409 | 0.293 | 0.642 |
| Laya | base checkpoint, published | 0.362 | 0.331 | 0.315 |
xMESH beats base Laya on this suite, but it does not beat Jev or the benchmark-specific Laya fine-tune. Accuracy by xMESH workflow was 0.266 agent-trace observability, 0.578 customer service, 0.492 invoice processing, and 0.574 security incidents.
A seeded option-order permutation changed 0 of 600 choice answers and produced zero probability drift because the Clef runtime canonicalizes option identifiers before inference.
On a short fixed state with identical three-option questions, Q8 llama.cpp on an Apple M4 measured 168 ms p50 for one question, 448 ms for five, and 847 ms for ten. Twenty and fifty questions exceeded this release's validated 2,048-token context; they are not advertised as supported.
Laya and Jev comparison rows above are figures published by the Laya benchmark page. They were not rerun in the xMESH environment.
Validated with Python 3.12 on an Apple M4 with 16 GB unified memory.
bash
python3.12 -m venv .venvsource .venv/bin/activatepip install -r requirements.txtpython examples/inference.py --request examples/request.json --device mpsUse --device cpu when MPS is unavailable. The request must contain state and questions; examples/request.json shows the typed contract.
bash
llama-server \ -m Singularity-xMESH-0.8B-BF16.gguf \ --host 127.0.0.1 --port 8081 \ -c 2048 -b 2048 -ub 2048 --parallel 1 -ngl 0 curl http://127.0.0.1:8081/v1/systemone \ -H 'Content-Type: application/json' \ --data-binary @examples/request.jsonxMESH is an advisory router. It must not receive wallet secrets, construct signed payments, bypass approval, invent endpoints, or execute arbitrary URLs. Runtime code must restrict choices to retrieved candidate IDs and enforce budgets, networks, schemas, and approval rules independently.
The model and Singularity runtime files are Apache-2.0. Qwen and Cloudflare Clef attribution is included in THIRD_PARTY_NOTICES.md.
| 35.63% |
| Executed cost regret | $0.00175 | $0.00000 | $0.00358 |
| Post-policy violations | 0 | — | — |
| 0.694 |
| Laya typed-decisions | trained on this benchmark's train split, published | 0.766 | 0.471 | 0.061 | 0.213 | 0.242 |
| Jev 1.13 | published | 0.727 | 0.580 | 0.148 | 0.144 | 0.391 |