TinyLM-16M deterministic smoke — q4_0

Observed TinyRustLM SLM1 q4_0 artifact from the 2026-06-25 implementation source snapshot. Deterministic runtime smoke only; no trained assistant quality is claimed.


Experimental

Verified 2026-06-25 00:00 UTC

Model details

Family
TinyRustLM deterministic smoke
Module type
Standalone specialist model
Skill identifier
tinylm16-q4
Capability evidence
Observed
Parameter count
17,048,064
Context size
512
Supported tasks
Runtime execution smoke, Parser and ABI regression
Formats
SLM1 custom .slm v1
Precision or quantization
q4_0
Reported file size
10,657,728 bytes (10.16 MiB) — measured from supplied artifact
Estimated runtime memory
Not measured. Artifact bytes exclude transfer duplication, decoded tensors, 8 MiB KV cache for TinyLM-16M, scratch, logits, and browser overhead.
Target environments
Browser WebAssembly CPU — implementation source and artifact present, Node WebAssembly ABI — locally rechecked only for TinyLM-16M q8_0
Backend compatibility
TinyRustLM scalar Rust/WASM runtime 0.1.0 — Experimental
License
Not separately declared; workspace metadata: Proprietary
Exact base requirement
Standalone SLM1 transformer artifact; no external base model at runtime.
Tokenizer requirement
Tokenizer is embedded in the artifact as BTOK or BPE1 and must match header vocabulary.
Module dependencies
tinyrustlm-runtime 0.1.0, matching SLM1 tokenizer/tensor contract
Artifact SHA-256
5ac2bd957113732be9ab2aadfa2d2efd54c3131e7e6c97be424fac6e72bb365b

Module contract

Input contract
UTF-8 prompt transferred to the synchronous raw WASM ABI; context must remain below the artifact maximum after tokenization.
Output contract
Generated UTF-8 text plus runtime diagnostics JSON. Model output has no execution authority.
Authority and side-effect scope
May generate local text inside the TinyRustLM runtime. No network, file, browser-action, tool, or write authority is granted.

Known limitations

Direct packed q4 matrix-vector execution with per-block scales; runtime-smoke only. Deterministic smoke weights; no trained assistant quality; no GGUF compatibility; no GPU; no modular routing.

Source and verification

Verification current

Primary source
Not recorded
Source revision
GGUF.MiRust.com-20260625.zip SHA-256 bd47fe8e91db5f5e5674ae4d77520b44248ab52117b170b345e1a0c87e636629; artifact SHA-256 5ac2bd957113732be9ab2aadfa2d2efd54c3131e7e6c97be424fac6e72bb365b
Reviewer
MiRust implementation source audit
Last verified
2026-06-25 00:00 UTC

Verification notes

Artifact path, byte length, SHA-256, SLM manifest, and source implementation were reviewed from GGUF.MiRust.com-20260625.zip (bd47fe8e91db5f5e5674ae4d77520b44248ab52117b170b345e1a0c87e636629).

This is an exact local artifact record from the inspected GGUF.MiRust.com source package.

Observed purpose

Direct packed q4 matrix-vector execution with per-block scales; runtime-smoke only.

Evidence boundary

The artifact exists, its byte length and SHA-256 were verified, and its sidecar admits it only for runtime execution smoke. It is not a downloadable asset from MiRust.com and is not presented as a useful trained assistant.

Runtime contract

Format: custom .slm v1. Backend: TinyRustLM scalar Rust/WASM CPU path. Tokenizer and exact shape are bound by the source artifact and manifest.

Known limitations

No independent task quality, safety, latency, peak-memory, cross-browser, device-compatibility, or license-rights conclusion is made. The project workspace declares Proprietary licensing; the individual model manifest does not provide a separate license identifier.