Proairos/Edge0-8B-A1B-preview-AideMemo-GGUF

Edge0 8B A1B Preview — AideMemo Recover-LoRA GGUF

An experimental Q4_K_M GGUF conversion of Edge0/Edge0-8B-A1B-preview, prepared and measured for AideMemo's local text-model evaluation.

This is a community derivative, not an official Edge0 release or its streaming engine. Recover-LoRA is merged into the weights, followed by requantization. Inference uses the original exact router without the prerouter. The upstream claim of roughly 1 GiB active memory does not describe this GGUF implementation.

Artifact

Field	Value
File	edge0-recover-q4_k_m.gguf
Bytes	4,823,894,848
SHA256	e9f9d7f9c96f04fb1dcc4017eb7e78ddcfd6c65cdba6059c0f252b672ead5edc
Source revision	cadc35c7059469fad8569a63ce6c851172bbca61
Tested runtime	llama.cpp ec91ab5add06555970f98d9c5361d884f3f530f8
Tested context	4,096 tokens
Canonical stop token	156895

The file includes the tokenizer and chat template. Consumers should preserve the native template and the explicit canonical stop token. Do not substitute a similarly named original MLX checkpoint or a different quantization and assume identical results. The original model's advertised context limit was not validated here.

Conversion

The published quantized source was reconstructed to F16 matrices and F32 vectors, Recover-LoRA was merged into 153 target tensors at scale 2.0, and the result was exported and requantized to Q4_K_M. The conversion excludes 66 prerouter/pregate tensors (61,210,624 bytes). Exact original routing remains active. This process is not bit-exact with the original Edge0 engine or full-precision model. The base lineage identified by Edge0 is inclusionAI/Ling-3.0-tiny-base.

Measured iPhone results

Physical iPhone 17 Pro, A19 Pro, approximately 12 GB RAM, iOS 27.2. Frozen September 2026 evaluation using a separate native benchmark app. Chinese / English pairs are shown in that order.

Metric	Chinese	English
Application score / 100	63.33	64.38
Strict public-question accuracy	75.00%	67.50%
Median short-prompt decode, tokens/s	49.772	53.017
Median inference-only first token, seconds	0.105	0.118

Application evaluation contains 24 tasks per language (90% semantic checks, 10% format). The public-question subset contains 40 questions per language with strict single-letter answers and different questions across languages. These are small diagnostic sets, not full MMLU/CMMLU scores or a universal ranking.

The common sustained workload completed in 300.53 active seconds. Timing-phase peak OS resident memory was 5.114 GB; Metal allocations were 4.949 GB. Those counters overlap and must not be added. They are observations, not minimum-RAM guarantees. First-token measurements exclude tokenization/context setup. Runs were sequential, not randomized thermal/energy comparisons.

Production-provider optimization check

A separate September 24 iPhone 17 Pro experiment used AideMemo's Swift provider with this exact file. Twenty interleaved synthetic requests compared fresh 4,096-token contexts with cleared, reusable 2,048-token contexts. All four marker A-B-A checks passed and all eight summary outputs matched. For the summary subset (four requests per policy), median first-token latency was 343 ms versus 318 ms; full generation was 1.757 s versus 1.755 s, so there was no meaningful end-to-end speed gain. Sampled physical-footprint peaks were 280.7 MB versus 245.5 MB, while resident memory including mapped weights remained about 5.2 GB. These are different, overlapping memory counters, not evidence of a sub-GB total model footprint. All requests reported nominal thermal state. A separate test without extended virtual addressing passed cancellation, unload-during-generation and recovery. These small synthetic checks do not replace the historical quality evaluation or establish battery savings, long-run thermal gains, or support for smaller-memory phones.

A later five-minute continuous run on the same iPhone completed 119 identical 45-token synthetic summaries, with one expected cancellation at the time limit. Median complete generation increased from 1.978 s in the first minute to 2.779 s in the last; generated tokens per decode second fell from 25.21 to 18.26. The OS thermal state reached fair, but never serious or critical. This is one observed sustained workload, not a controlled energy or thermal-policy comparison. Separate load profiling observed 11.046 s for the first load in a process and about 2 s for subsequent fully verified reloads; OS file-cache state was uncontrolled.

Bounded application-task quality follow-up

A separate Mac experiment ran 12 frozen synthetic bilingual journal/profile cases with the production prompts and schemas, and compared one candidate instruction suffix (24 total requests). Baseline exact schema conformance was 2/12; the candidate reached 5/12 but retained semantic errors, including negation and personal-fact ownership mistakes, and failed the held-out promotion gate. The candidate was not adopted. These challenging task-shaped fixtures are not a representative accuracy benchmark and do not replace the historical score above. This version remains experimental for complex structured memory extraction.

A subsequent frozen comparison used 18 synthetic cases (12 exposed regressions and six fresh cases), with 36 requests and unchanged prompts/weights. Schema-specific decoding improved exact-schema conformance from 8/18 to 17/18, but accepted outputs with manually identified semantic/task issues increased from 4/7 to 8/10. One accepted profile stored a neighbor's phone number as the user's phone. The schema decoder was not enabled in the App. These judgments were nonblind, with cases divided between two reviewers; they are not a population accuracy estimate.

Weight versus application changes

The downloadable weight file has not changed during the runtime optimizations. Adaptive context reuse, staged loading feedback, literal handling of native control markers, numeric/deadline checks and a within-draft birthday ambiguity guard are AideMemo application/runtime features; downloading this GGUF alone does not enable them in other clients. The birthday guard can reject different representations of the same date and does not establish fact ownership. In offline replay it changed no acceptance decisions in the 36-request comparison because the relevant birthday outputs were already rejected for other reasons. No overall factual-accuracy, battery-efficiency or meaningful full-response speed improvement is claimed.

Limitations

• Fluent summaries can invent details, including a deadline absent from the source. Arithmetic, filtering and multi-condition instructions can fail. Preserve source grounding and review generated memory summaries.
• Four reviewed writing samples cannot establish a general hallucination rate.
• This artifact does not reproduce the official streaming/offload engine, prerouter, or official memory footprint. Full logits equivalence is not established.
• AideMemo's production GGUF provider separately passed bounded Mac and iPhone load, structured-JSON generation, canonical-stop and lifecycle checks with this exact file. These are not a complete production phone quality rerun.
• Larger or continuous workloads and lower-memory phones require their own validation.

License and attribution

The pinned upstream model card declares Apache-2.0. The included LICENSE preserves the Apache 2.0 text from the pinned Edge0 source tree, and NOTICE identifies this derivative and its modifications. Original models and names belong to their respective authors; this publication does not imply endorsement.

Source engine: Edge0-AI/Edge0. Underlying runtime: llama.cpp.