{ "schema_version": 1, "builder_plan": { "scenario": "A self-hosted assistant is too expensive because its model uses most of the GPU memory.", "change": "Try the paper's supported lower-precision weights on a copy of your serving setup.", "baseline": "The same model at its original precision, on the same GPU.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Peak memory, task pass rate, latency, and total serving cost at the same context and load.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2609.29397", "title": "Baseline Shape Decides the Verdict: A Controlled Re-Examination of Ternary Language Models at 60K Parameters", "summary": "The examined method replaces full-precision weights with 1.58-bit ternary weights and a full-precision-then-ternary training schedule for microcontroller-class models, but this re-examination primarily replaces the prior single-seed evidence with a multi-seed controlled comparison. Costs include a measurable quantization penalty that varies strongly by architecture (+5.3% for the best transformer versus +19.5% routed and +28.1% gated SSM) plus a fragile two-stage schedule that requires a stage-2 learning rate roughly 10x the pretraining peak to beat all-ternary training. It can fail if the baseline shape is untuned, since depth/width choice alone spans 22.6% in validation loss and shape ordering reverses with budget, meaning an apparently superior architecture may only reflect a poorly shaped baseline.", "rationale": "The findings are directly usable at this reader's scale (98 runs on one laptop) but only as cautionary guidance: they show ternary sub-1M claims are baseline- and schedule-sensitive, so adoption is justified only after workload-specific validation of baseline shape and fine-tuning learning rate.", "technique": "Ternary 1.58-bit weights for sub-1M models", "family": "quantizacao", "recommendation": "testar", "published": "2026-09-24", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.29397", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "The examined method replaces full-precision weights with 1.58-bit ternary weights and a full-precision-then-ternary training schedule for microcontroller-class models, but this re-examination primarily replaces the prior single-seed evidence with a multi-seed controlled comparison. Costs include a measurable quantization penalty that varies strongly by architecture (+5.3% for the best transformer versus +19.5% routed and +28.1% gated SSM) plus a fragile two-stage schedule that requires a stage-2 learning rate roughly 10x the pretraining peak to beat all-ternary training. It can fail if the baseline shape is untuned, since depth/width choice alone spans 22.6% in validation loss and shape ordering reverses with budget, meaning an apparently superior architecture may only reflect a poorly shaped baseline.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.29397", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "At a 130M-byte training budget the routed ternary block beats parameter-matched transformers by 22.2-24.0% in validation loss, but a plain gated diagonal SSM block beats the routed model by a further 9.1%; no multiplicative factor is reported.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.29397", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [], "equations_status": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }