Researchers increasingly need transcripts that are both high-quality and private. Epilude’s July 14, 2026 release of Model 4 improves local cleanup for Mac dictation (1.5 GB on-device, offline, and with critical meaning errors down 56%) and shows why on-device transcription for qualitative research matters. In this post we summarize the changes, explain what they mean for interview- and diary-based studies, and map a concrete workflow to turn those private transcripts into thematic, cross-segment, and visual analyses without ever sending raw audio to the cloud.
Key Takeaways
Evidano is an AI-powered qualitative data analysis platform that ingests private on-device transcripts and produces thematic and visual insights without sending data to third parties. Epilude announced Model 4 on July 14, 2026, a local cleanup model that runs on Apple Silicon, installs as a 1.5 GB package, and reduced critical meaning errors by 56% (from 9 to 4) on Epilude’s internal 90-scenario benchmark.
- Model 4 runs offline on Apple Silicon Macs, preserving audio and transcript privacy while performing local cleanup.
- Model 4 reduced critical meaning errors by 56% (from 9 to 4) on Epilude’s internal 90-scenario benchmark, improving trust in automated transcripts.
- Full cleanup with Model 4 requires Apple Silicon and 16 GB of RAM, otherwise expect basic formatting and offline transcription only.
- Evidano ingests cleaned, on-device transcripts and performs thematic, cross-segment, and visual analyses without sending data to third-party models.
Fast take: what changed (source)
Epilude announced Model 4 on July 14, 2026.
Model 4 is a new on-device cleanup model that runs locally in Epilude’s Local Mode and aims to make dictation transcripts cleaner without leaving a Mac. Read the original post at Epilude.
- On-device cleanup model (Model 4) built on the Qwen family, same 1.5 GB footprint and speed.
- Critical meaning errors down 56% (Model 3: 9 → Model 4: 4) on Epilude’s internal 90-scenario benchmark.
- Full cleanup requires Apple Silicon Macs with 16 GB RAM; otherwise transcription plus basic formatting runs offline.
Findings snapshot (key metrics)
| Metric | Model 3 | Model 4 | Notes / Implication |
|---|---|---|---|
| Scenarios passed (of 90) | 69 | 78 | Improved robustness across punctuation, self-corrections, mixed language |
| Critical meaning errors | 9 | 4 | 56% reduction, fewer outputs that change meaning |
| Download size | 1.5 GB | 1.5 GB | One-time, local install |
| Runtime | Same | Same | Greedy, deterministic decoding on Apple Silicon |
| Connectivity | N/A | Offline | Audio/text never leaves the Mac |
How Model 4 changes the transcript pipeline
Model 4 is a cleanup model that takes speech-to-text output and sequences it into sendable writing, fixing punctuation, removing filler, resolving false starts, and preserving speaker tone or mixed-language phrases.
Epilude’s evaluation focuses on zero-tolerance critical errors (meaning changes, instruction leaks); Model 4 passes nine more benchmark scenarios than Model 3 and reduces critical errors from 9 to 4.
- Deterministic decoding: outputs are stable across runs, which matters for reproducible coding.
- Privacy-first: everything runs locally on Apple Silicon, useful when consent or regulation prohibits cloud audio processing.
- Known limits: long, rambling inputs still trigger most residual failures; cloud models still outperform on some long-input metrics.
So what for qualitative researchers and UX teams
Trustworthy transcripts = faster coding
A 56% drop in critical meaning errors means fewer manual corrections for date/name mistakes and self-corrections, which reduces the time spent on transcription quality control.
Deterministic outputs reduce inter-run variability, making AI-assisted coding and automated theme extraction more reliable.
Privacy & compliance
Processing entirely offline aligns with IRB and data protection constraints where audio cannot be uploaded, which simplifies deployment in regulated environments.
Local Mode’s on-device cleanup lets teams maintain raw audio and transcripts within institutional environments, lowering legal and ethical friction for deploying AI-assisted workflows.
Multilingual & tone-preserving transcripts
Model 4 preserves mixed-language segments instead of silently translating them and retains casual speech markers when requested, which is useful for discourse analysts and UX researchers preserving speaker voice.
Preserving speaker voice and mixed-language phrases helps analysts keep context for thematic interpretation and exemplar quote selection.
Do more, faster with Evidano (map to your workflow)
Problem: Private, messy transcripts
You have high-quality, offline transcripts from Model 4 but still need thematic analysis, segment comparisons, and stakeholder-ready visuals.
Private, messy transcripts require import, normalization, coding, clustering, and visualization before they become publishable insights.
Solution: Import + analyze in Evidano
Evidano ingests on-device transcripts (text files, time-stamped transcripts) and runs thematic, content-frequency, and cross-segment analyses without sending data to third-party models.
Key Evidano features that map to this use case include transcription import and normalization, codebook import and AI-assisted coding, thematic clustering, cross-segment frequency analysis, and visualizations such as word clouds, co-occurrence networks, and hierarchical code trees.
Security note: Evidano encrypts data and does not use customer data to train external models, keeping the privacy gains from on-device cleanup end-to-end.
Quick wins you can expect
Reduce manual QC time using automated spot-checks against Model 4’s deterministic outputs.
Faster synthesis through auto-generated themes and exemplar quotes you can drag into stakeholder memos.
Segmented insights: compare response patterns by cohort, such as region or persona, with frequency and co-occurrence metrics.
Checklist: From Model 4 transcripts to publishable insights
Follow these steps to turn private, on-device transcripts into research outputs.
- 1) Export cleaned transcripts from Local Mode (Model 4) as UTF-8 text or time-stamped JSON.
- 2) Upload files to Evidano via secure, encrypted import and set a custom dictionary for names and terms (Evidano).
- 3) Run automatic cleanup and PII redaction if required, then import or build a codebook.
- 4) Use AI-assisted coding to apply codes at scale and review edge cases flagged by the platform.
- 5) Generate thematic summaries, cross-segment frequency tables, and co-occurrence visuals.
- 6) Export quotes and visuals into a stakeholder slide deck or decision memo.
Limitations & research guardrails
Model 4 improves many failure cases but still struggles with long, rambling passages, so plan a QA pass for long monologues or narrative interviews.
For highly sensitive clinical or legal work, retain human review and explicit consent; Evidano supports PII redaction and access controls to help with this.
- Model 4 requires Apple Silicon and 16 GB RAM for full cleanup; for smaller machines expect basic formatting only.
- Epilude’s benchmark is internal and not directly comparable to external metrics, so validate on a held-out sample from your corpus before full automation.
FAQ: Private On-Device Transcription
What is Model 4 and when was it released?
Model 4 is an on-device cleanup model released by Epilude on July 14, 2026.
Epilude positioned Model 4 as a local cleanup model for Mac dictation that post-processes speech-to-text output to improve punctuation, remove filler, and preserve speaker tone.
How much did Model 4 reduce critical meaning errors?
Model 4 reduced critical meaning errors by 56% on Epilude’s internal benchmark, dropping from 9 to 4 critical errors across 90 scenarios.
Epilude’s internal 90-scenario benchmark reports fewer outputs that change meaning, which improves transcript trustworthiness for qualitative coding.
What hardware is required to run Model 4 fully?
Full cleanup with Model 4 requires Apple Silicon Macs with 16 GB of RAM.
On machines with less memory Model 4 still provides offline transcription and basic formatting but not full cleanup performance.
How do researchers use Model 4 transcripts in a secure analysis pipeline?
Researchers can export cleaned, local transcripts and import them into Evidano to run thematic and cross-segment analyses without sending data to third-party models.
Evidano ingests text and time-stamped transcripts, supports PII redaction, and performs AI-assisted coding and visualization while encrypting customer data.
Conclusion; Ready to run private transcripts at scale?
Epilude’s Model 4 demonstrates that high-quality cleanup and strict privacy are not mutually exclusive.
For qualitative teams, that means transcripts you can trust and analyze without moving audio off-device; to turn local transcripts into reproducible themes, segment comparisons, and stakeholder-ready visuals, Try Evidano for free.