Follow-Up Research · June 2026 · Continuation of Part 1 105 agents · 23 sources (all primary) · 108 claims extracted

Follow-Up Research · June 2026 · Continuation of Part 1

Answering the Hard Questions About Local LLMs

A second adversarially-verified research pass targeting only primary sources — peer-reviewed papers, arXiv preprints with auditable methodology, and official documentation — in response to the four open questions left unanswered in Part 1.

Research agents
105
Sources (all primary)
23
Claims extracted
108
Claims verified
6
Claims killed
19
OVERVIEW

What this pass found — and didn't

This is a follow-up to Part 1. The first research pass confirmed one architectural fact and killed 24 claims from grey literature. This pass targeted the four unanswered questions using only arXiv preprints with explicit methodology, peer-reviewed papers, and official documentation. The result: 6 confirmed findings, 2 questions partially answered (Q1, Q3), and 2 questions that remain genuinely unanswered at the primary-source standard (Q2, Q4). The findings are detailed by question below.

Q1 Partial answer
Throughput benchmarks across GPU tiers
Q2 Still unanswered
On-premise vs. cloud cost break-even
Q3 Partial answer
Quantization quality effects, empirically
Q4 Still unanswered
Documented enterprise deployments
Q1 — PARTIAL ANSWER

What primary-source throughput benchmarks actually exist?

Two peer-reviewed or formally published studies were found and verified. Neither covers the specific consumer GPU tiers most practitioners care about (RTX 3060, 3090, 4090 under realistic concurrent workloads) — but they represent the best available primary-source evidence.

PowerInfer on RTX 4090 Verified — High
Vote 2–1 (scope limitation noted, not a fabrication)
Claim PowerInfer on a single RTX 4090 achieves 82% of an NVIDIA A100's token generation rate for OPT-30B inference at batch size 1 — the typical single-user local deployment scenario. Experimental setup: RTX 4090 (24GB) paired with an i9-13900K versus an 80GB A100 running vLLM, both at batch size 1, input length 1.
Source arXiv:2312.12456 · PowerInfer · Peer-reviewed at SOSP 2024 (top-tier systems venue)
arxiv.org/abs/2312.12456
Caveat Critical scope: the 82% figure applies only at batch size 1. PowerInfer's activation-sparsity advantage degrades at larger batch sizes. This is a research prototype, not a production runtime, and results don't generalize to concurrent-user workloads.
Argonne LLM-Inference-Bench Verified — High
Vote 3–0 unanimous on scope; 2–1 on framework detail
Claim The Argonne LLM-Inference-Bench study is the most comprehensive primary-source benchmark available. It covers NVIDIA GPUs, AMD GPUs, Intel Habana, and SambaNova accelerators, using five inference frameworks — vLLM, llama.cpp, TensorRT-LLM, DeepSpeed-MII, Sambaflow — on LLaMA, Mistral, and Qwen models at 7B and 70B parameter scales. The benchmark repository is public and reproducible.
Source arXiv:2411.00136 · Argonne National Laboratory · Published at IEEE (October 2024)
arxiv.org/abs/2411.00136 · github.com/argonne-lcf/LLM-Inference-Bench
Caveat The paper focuses on enterprise-class hardware (A100, H100, MI300X). Consumer GPU tier data (RTX 3060, 3090, 4090) is not the study's emphasis. It remains the best publicly verifiable multi-framework comparison, and the GitHub repository allows reproducing the benchmarks on any supported hardware.
Gap — What remains missing for Q1

No peer-reviewed paper providing reproducible throughput benchmarks specifically for consumer GPU tiers (RTX 3060, 3090, 4090) under realistic multi-user concurrent workloads was found and verified. The XDA Developers paper (arXiv:2410.04466, "Large Language Model Inference Acceleration: A Comprehensive Hardware Perspective") was fetched and processed but yielded claims that failed 1–2 verification — the specific RTX 3090 and M2-Ultra figures could not be independently corroborated. The Argonne repository is the recommended starting point for practitioners who need reproducible data.

Q2 — STILL UNANSWERED

On-premise vs. cloud cost break-even: the literature is broken

Five cost-model papers were found, fetched, and verified. All five failed adversarial verification at 0–3 votes. The one surviving cost-related finding reframes why break-even calculations are inherently difficult.

Gap — No cost model survived

Every specific cost break-even figure encountered in this research pass was killed. This includes two arXiv preprints (2509.18101 and 2601.09527) that presented auditable frameworks with explicit hardware prices and API rate assumptions. The figures could not be independently corroborated. The underlying methodological problem — that GPU utilization rate is either a fixed assumption or silent 100% in every public LLM cost calculator — was flagged by one paper (2606.11690) but that paper's own numerical claims also failed verification.

API price-performance decline Verified — Medium
Vote 2–1 (arXiv preprint, not yet peer-reviewed)
Claim Frontier LLM API price-performance drops approximately 5–10× per year, based on logit regression across 138 price/benchmark data points from April 2024 to November 2025. This rate is independently corroborated at the lower bound by Epoch AI's Cottier et al. 2025, which reports even higher rates on different benchmarks. The implication: any static break-even calculation has a shelf life measured in months, not years.
Source arXiv:2511.23455v2 · Gundlach, Lynch, Mertens, Thompson · November 2025
arxiv.org/pdf/2511.23455
Caveat This finding cuts both ways for on-premise economics. Falling API prices compress the cost advantage of local deployment over time. A break-even analysis from mid-2024 may overstate on-premise savings by 5–10× when evaluated twelve months later. Any practitioner building a business case for on-premise deployment must model API price-performance as a declining variable, not a fixed baseline.

Why no break-even paper survived

The three most specific papers (2509.18101, 2601.09527, 2606.11690) all present numerical claims that verification agents could not corroborate from independent sources. The most likely explanations: (1) figures are derived from pricing snapshots that change faster than the papers' update cycle, (2) hardware cost assumptions (e.g., "RTX 5090 at $2,000") rely on MSRP figures disputed by market pricing, and (3) GPU utilization assumptions baked into cost-per-token calculations are not independently auditable. The methodology gap identified by 2606.11690 — that utilization is systematically underspecified — appears real even if that paper's own numbers didn't survive.

Q3 — PARTIAL ANSWER

Quantization quality effects: what the peer-reviewed data shows

Two papers survived adversarial verification. Both are arXiv preprints without confirmed peer review, but both have explicit, auditable methodology and were verified directly against their published tables. Both apply to LLaMA-family models only — no verified data was found for Mistral, Phi, or Gemma.

GGUF quantization on Llama-3.1-8B Verified — Medium
Vote 2–1
Claim Among all GGUF quantization levels tested on Llama-3.1-8B-Instruct, Q3_K_S produces the largest average quality loss: 5.73% across five benchmarks (GSM8K, HellaSwag, IFEval, MMLU, TruthfulQA), including a 9.32-point drop on math reasoning (GSM8K: 68.31 vs. F16 baseline 77.63). It achieves the highest compression (77.23% size reduction) but at the most visible quality cost of any scheme above Q2. Q4_K_M sits in a different zone: substantially smaller but with moderate quality loss on instruction-following tasks specifically.
Source arXiv:2601.14277 · "Which Quantization Should I Use?" · January 2026
arxiv.org/abs/2601.14277
Caveat Single model only (Llama-3.1-8B-Instruct). The Q8_0 near-F16 parity claim — commonly accepted wisdom — received 0–3 votes from this same paper's data, meaning it could not be independently confirmed. Do not generalize these numbers to other model families or sizes without verification.
AWQ vs. OmniQuant on LLaMA3-8B Verified — Medium
Vote 2–1
Claim At 4-bit quantization, AWQ and OmniQuant perform comparably on LLaMA3-8B (AWQ: perplexity 7.98, accuracy 61.23%; OmniQuant: perplexity 8.82, accuracy 57.36%). At 3-bit, there is a sharp divergence: OmniQuant degrades severely (perplexity 17.53, accuracy 36.09%) while AWQ retains substantially better quality (perplexity 9.83, accuracy 54.65%). This "3-bit cliff" for OmniQuant on LLaMA3-8B is independently noted in a separate paper (GANQ, arXiv:2501.12956), which reports OmniQuant cannot quantize LLaMA3-8B on a single RTX 4090 in some 3-bit configurations.
Source arXiv:2502.13178 · Systematic PTQ benchmarking study · February 2025 (revised May 2025)
arxiv.org/abs/2502.13178
Caveat LLaMA-family specific. Cross-architecture results in the same paper show divergence at 4-bit for non-LLaMA architectures. The 2-bit collapse finding (0–3 votes) means no verified data exists for the very-low-bit tier. The paper is an arXiv preprint not yet confirmed peer-reviewed.

What the data as a whole says about the Q4–Q8–F16 hierarchy

Taken together with the Part 1 confirmed finding about GGUF architecture, the verified evidence points to: Q3 and below produce measurable quality loss on reasoning tasks; Q4 is the practical trade-off point where compression is substantial and quality loss is moderate; the 4-bit tier is method-dependent (AWQ vs. OmniQuant diverges at 3-bit); and the Q8 ≈ F16 claim — perhaps the most repeated assertion in local LLM discussions — could not be confirmed from any primary source in either research pass.

Gap — What remains missing for Q3

No verified empirical data covers quantization effects on Mistral, Phi, or Gemma model families at any level. The two surviving findings apply exclusively to LLaMA-family models. Practitioners deploying Mistral or Phi variants locally have no peer-reviewed primary-source data on quality degradation at their target quantization level.

Q4 — STILL UNANSWERED

Documented enterprise deployments: a genuine literature gap

This was the most thoroughly searched angle and returned zero verified claims. The search covered EMNLP 2024 and 2025 industry tracks, NeurIPS proceedings, arXiv enterprise deployment papers, and company-authored technical reports. The result is not a search failure — it reflects a genuine gap in the primary literature.

Gap — Zero claims survived for Q4

No enterprise deployment of a local LLM with a verifiable organizational name, specific use case, and measurable outcomes was confirmed. Several sources came close: the WhatsCode paper (arXiv:2512.05314) describes a named deployment at WhatsApp over 25 months generating 3,000+ accepted code changes — but it is a cloud-based system, not a local LLM deployment. A claim about SkillsFuture Singapore achieving 62.5x faster case processing using GraphRAG with a local model was extracted from an EMNLP industry track source but failed adversarial verification (could not independently confirm the 62.5x figure or the "local" specificity). A "regulator-approved retail banking AI" appeared in one paper but the organizational name was withheld.

Why this question remains unanswered

There are structural reasons why primary-source evidence for enterprise local LLM deployments is sparse. Organizations in regulated industries — the most likely adopters of local deployment for compliance reasons — have strong incentives not to disclose their AI infrastructure. Conference paper authorship in industry tracks often anonymizes or omits the deploying organization. Vendor case studies (which do name organizations) are systematically grey literature and do not meet a primary-source standard. The gap is real.

Note — Where to look instead

Practitioners seeking enterprise deployment examples should search: EMNLP and ACL industry tracks (2024–2025), procurement notices on SAM.gov and similar government portals for public-sector AI infrastructure, and corporate SEC filings and earnings calls for named AI deployment disclosures. These sources are below the primary-source threshold used here but are more likely to contain named organizational data than academic preprints.

KILLED CLAIMS

Interesting near-misses and what failed verification

The following claims were extracted from primary or near-primary sources but failed adversarial verification. Several are directionally plausible — they failed because independent corroboration was not located within the verification pass, not necessarily because they are false.

VoteClaimSource
1–2RTX 4090 + AWQ INT4 increases Llama-2-7B from 52 t/s to 194 t/s — the only specific consumer-GPU throughput figure found in a primary sourcearXiv:2410.04466
1–2RTX 3090 running Llama2-7B with mixed 2/4-bit quantization achieves 45.2 t/s — the only RTX 3090 figure found in a primary sourcearXiv:2410.04466
0–3Apple M2-Ultra running Llama2-7B at 4-bit via llama.cpp achieves 6 t/s single-core and 32 t/s with eight coresarXiv:2410.04466
0–3Q8_0 achieves near-identical downstream task performance to F16 across MMLU, HellaSwag, TruthfulQA, and GSM8K (widely-held belief, failed from this paper's own data)arXiv:2601.14277
0–3Q4_K_M causes a 6.58-point IFEval drop vs. F16 — instruction-following disproportionately sensitive to 4-bit quantizationarXiv:2601.14277
0–3Sub-30B on-premise deployment breaks even vs. commercial APIs in 0.3–3 monthsarXiv:2509.18101
0–3RTX 5090 delivers 411 t/s for a RAG-8k workload on Qwen3-8B NVFP4 at concurrency 8arXiv:2601.09527
1–2Every public LLM cost calculator treats GPU utilization as fixed input or silent 100% assumption — identified as the critical methodological gap in cost modelingarXiv:2606.11690
0–32-bit quantization causes complete language-capability collapse regardless of model size (AWQ on LLaMA-65B: perplexity 7.4×10⁴)arXiv:2502.13178
STILL OPEN

Refined open questions after both research passes

Two passes of adversarial research have narrowed the unknowns. The questions below are now sharply scoped — they are not general inquiries but specific gaps in the primary literature that practitioners should be aware of when making infrastructure decisions.

R1   What reproducible, peer-reviewed throughput benchmarks exist for consumer GPU tiers (RTX 3060, 3090, 4090) under realistic multi-user concurrent workloads? The two surviving Q1 findings (PowerInfer, Argonne) cover batch-size-1 and enterprise hardware respectively. The consumer-GPU concurrent-request scenario — the most common real-world deployment pattern — has no verified primary-source benchmark.

R2   Does the 3-bit quantization cliff observed for OmniQuant on LLaMA3-8B generalize to Mistral, Phi, and Gemma families? Every verified quantization finding in both passes applies exclusively to LLaMA-family models. Practitioners deploying other architectures have no primary-source quality degradation data at any quantization level.

R3   Given that cloud API price-performance drops 5–10× per year, what utilization rate and workload volume make on-premise deployment economically rational — and what methodology produces a utilization-adjusted break-even that survives adversarial verification? All five cost-model preprints failed; the utilization-underspecification critique (arXiv:2606.11690) appears valid but its own numbers also failed.

R4   What primary-source evidence exists for named enterprise local LLM deployments? The EMNLP and NeurIPS industry tracks were searched and produced zero verified claims meeting the standard. This is a structural gap: regulated-industry adopters have the strongest incentives to deploy locally and the strongest reasons not to publish about it.

Key Sources

arXiv:2312.12456 — PowerInfer / SOSP 2024 arXiv:2411.00136 — Argonne LLM-Inference-Bench / IEEE github.com/argonne-lcf/LLM-Inference-Bench arXiv:2511.23455 — API price-performance arXiv:2601.14277 — GGUF quantization on Llama-3.1-8B arXiv:2502.13178 — PTQ systematic study, AWQ vs. OmniQuant arXiv:2410.04466 — LLM Inference Acceleration: A Comprehensive Hardware Perspective (killed claims) arXiv:2509.18101 — on-premise break-even cost model (killed) arXiv:2601.09527 — RTX 5090 RAG workload throughput (killed) arXiv:2606.11690 — GPU utilization methodology gap (killed) arXiv:2501.12956 — GANQ (OmniQuant 3-bit cliff corroboration) Part 1: Local LLM Applications & Ecosystem