Can Quantized Local LLMs Do Qualitative Coding? A 2026 Test

Running an open-weight model on your own machine is the usual answer when interview data cannot leave the building. The catch is hardware. A full-precision model needs a large GPU, so most local setups use a quantized model: the same weights stored with fewer bits, which saves memory and runs faster. A new paper by Aisvarya Adeseye, Jouni Isoaho and Adeyemi Adeseye, accepted to the Intelligent Systems Conference (IntelliSys) 2026, measures what that compression costs in thematic coding, and whether a simple verification loop can recover it.

What did the study test?

The study tested how far quantization degrades theme extraction, frequency counting and hallucination in qualitative analysis, using real interviews rather than generic reasoning benchmarks.

• Data: 82 semi-structured interviews, each 45–60 minutes, producing transcripts of 8,000 to 13,000 words. The topic was privacy concerns about gamification inside organizations, with participants from NGOs, companies, universities, government agencies and healthcare.
• Experts vs non-experts: 33 participants who reported moderate or high familiarity with privacy and data protection were classed as experts; the other 49 as non-experts. The split lets the authors compare precise technical language with everyday, vaguer descriptions.
• Model: LLaMA-3.1 (8B), run locally through llama.cpp with no external APIs, at BF16 (full precision) and at 8-, 4-, 3- and 2-bit precision.
• Quantization methods: SmoothQuant, INT8 and GGUF Q8_0 (8-bit); Q4_K_M, NF4 and GPTQ (4-bit); HQQ and AWQ (3-bit); SPQR and GGUF Q2_K (2-bit).
• Ground truth: Two researchers coded all transcripts in NVivo (Cohen's kappa 0.87 for theme identification, 89.4% overall agreement). Their coding was merged with manually corrected BF16 model output, keeping only findings both researchers verified.

Outputs were scored on eight metrics, including theme F1, semantic drift, hallucination rate, theme consistency across five repeated runs, keyword frequency correlation and clustering agreement (Adjusted Rand Index).

How does the multi-pass verification method work?

The method splits the job into small, checked steps, so a weak model never has to do everything in one long prompt.

1. Extract. A short, strict prompt ("use only the transcript, do not guess") asks for themes, sub-themes and exact supporting quotes, returned as JSON.
2. Verify themes. A second prompt checks every theme and quote against the transcript and deletes anything unsupported. If it finds errors, the step repeats.
3. Count. Using only the verified themes, the model counts how often each theme and sub-theme appears.
4. Verify counts. A final prompt corrects the counts without adding new themes, and only then does the next transcript begin.

Verification stopped when no further unsupported items were found. According to Adeseye et al. (2026), 93% of transcripts converged within two verification cycles, and none needed more than three passes.

How accurate are quantized LLMs at thematic analysis?

At 8 and 4 bits, accuracy was close to full precision after verification; at 3 and 2 bits it was not. The gap was widest on non-expert transcripts, where informal wording made errors more likely for every model, and for the human coders too.

The table shows theme F1 (higher is better) and hallucination rate (lower is better) on non-expert transcripts, before and after verification, for one representative method per precision level. Figures are from Table 1 of the paper.

Model (precision)	Theme F1, before	Theme F1, after	Hallucination, before	Hallucination, after
Human coders (reference)	0.80	–	0.00	–
BF16 (full precision)	0.65	0.82	0.22	0.09
SmoothQuant (8-bit)	0.63	0.81	0.24	0.10
NF4 (4-bit)	0.61	0.79	0.25	0.11
HQQ (3-bit)	0.37	0.58	0.40	0.22
SPQR (2-bit)	0.25	0.45	0.50	0.30

On expert transcripts the starting point was much higher: theme F1 before verification was 0.90 for BF16, 0.88 for SmoothQuant, 0.86 for NF4, 0.62 for HQQ and 0.50 for SPQR. Clear terminology helped every model, and the effect of compression grew as precision fell.

Does the quantization method matter, not just the bit level?

Yes. Models at the same bit level behaved differently. Among 4-bit options, NF4 reached a non-expert theme F1 of 0.61 before verification, against 0.58 for Q4_K_M and 0.55 for GPTQ. Among 8-bit options, SmoothQuant (0.63) beat GGUF Q8_0 (0.58). Choosing a format from a download menu is a methodological decision, not a technical detail.

Which part of the verification helps most?

Theme-level verification did most of the work. The authors report that the first verification pass accounted for about 71% of the total reduction in hallucination; the second pass mainly improved frequency alignment (+6.4% Pearson correlation) and consistency across runs (+4.1%). For HQQ (3-bit), non-expert hallucination fell from 0.40 to 0.22, a 45% reduction. Frequency checks mattered for counts: for Q4_K_M, non-expert frequency correlation rose from 0.61 to 0.79.

What does verification cost?

Verification added an average of 34.7% to runtime. A single pass took 41.8 seconds per transcript at 8-bit and 38.6 seconds at 4-bit, with each verification cycle adding about 17.2 seconds. Memory use stayed flat at roughly 14.2 GB of VRAM for 8-bit and 9.6 GB for 4-bit, on a single RTX 4090.

What this means for researchers

The practical message is that verification matters more than raw model size, with some caveats worth reading before you rely on it.

• Do not go below 4 bits for coding. Even with verification, 3-bit and 2-bit models stayed far below the human reference.
• Build a check step into any AI coding workflow. The same verification also lifted full-precision BF16, from 0.65 to 0.82 theme F1 on non-expert transcripts. Asking a model to defend each theme with an exact quote is cheap insurance at any precision.
• Expect more errors with lay participants. Consumer, patient and community interviews resemble the non-expert group here, which was harder for every model and for human coders.
• Read "after" numbers critically. A hallucination rate of 0.10 still means about one statement in ten is unsupported. Human review of the final themes remains necessary.

Some limits are worth noting. Only one model, LLaMA-3.1 (8B), was tested, on interviews from a single domain, and the authors call for tests on Gemma, Mistral and Phi. The ground truth also drew on corrected BF16 output; the authors state that every candidate theme was re-validated by the human coders, but a benchmark built partly with the same model family may flatter it. These points do not undermine the direction of the findings, but they limit how far the exact numbers transfer.

If your data can be processed by a hosted tool, ThemeLens applies the same principle of grounding each theme in quoted evidence that you can trace back to the transcript. If you still need to collect the interviews, the AI Interviewer produces transcripts ready for that kind of analysis.

FAQ

What is a quantized LLM?

A quantized LLM stores its weights with fewer bits, for example 4 bits instead of 16. It needs less memory and runs faster, which makes local, private analysis possible on ordinary hardware, but it can lose accuracy.

Can a 4-bit model be used for qualitative coding?

In this study, 4-bit LLaMA-3.1 (8B) became competitive with 8-bit models once multi-pass verification was added: NF4 reached a non-expert theme F1 of 0.79, against 0.81 for SmoothQuant 8-bit. Without verification it was noticeably weaker.

Why do LLMs hallucinate more on non-expert interviews?

Non-experts used informal, ambiguous wording, so the intended meaning was often indirect. The authors found this raised error and hallucination rates for all models, and increased disagreement between human coders as well.

Does verification make AI coding as good as human coding?

Not quite. After verification, BF16 and 8-bit models matched or came close to the human reference on theme F1, but they still produced some unsupported statements, while the human coders produced none.

Source: Adeseye, A., Isoaho, J., & Adeseye, A. (2026). Improving Quantized Model Performance in Qualitative Analysis with Multi-Pass Prompt Verification. arXiv:2605.20193. Accepted to the 12th Intelligent Systems Conference (IntelliSys 2026).

Last updated: November 3, 2026

This article is an independent editorial summary of third-party research. QualiTaTi is not affiliated with the authors or the conference.