9. Alternative Local Models
9.1 Candidate model families
Several open-weight model families can perform the same document-to-structure task. Their suitability depends on local memory, latency, multilingual coverage, instruction following, long-context behavior, and structured-output reliability. The table summarizes plausible alternatives as of July 2026; it does not claim that any alternative will outperform the evaluated Gemma configuration without a controlled benchmark.
| Model | Scale/context | Multilingual scope | Relevance to this task |
|---|---|---|---|
| Gemma 4 26B A4B IT | 25.2B total / 3.8B active; up to 256K | 140+ languages | Evaluated baseline; strong multilingual breadth, MoE efficiency, official thinking toggle and QAT variants [4-6]. |
| Qwen3 30B A3B | 30.5B total / 3.3B active; 128K with YaRN | 100+ languages and dialects | Close architectural alternative with thinking/non-thinking modes and strong multilingual instruction following [8]. |
| Mistral Small 3.2 24B | 24B dense; 128K | Dozens of languages; narrower than Gemma | Strong precise instruction following and robust function calling; local deployment is feasible when quantized [9,10]. |
| Llama 4 Scout | 17B active MoE; very long context | 12 officially supported languages; pretraining includes more | Capable open-weight alternative, but official multilingual support is narrower than Gemma 4 [11]. |
| Ministral 3 14B Instruct | 14B dense; local/edge oriented | Model-specific evaluation required | Lower memory and potentially higher throughput; may trade extraction coverage for speed [12]. |
| Phi-4 mini instruct | 3.8B; 128K | Broad multilingual use claimed, but smaller capacity | Useful latency-oriented baseline; should be expected to require task-specific testing and possibly fine-tuning [13]. |
9.2 Multilingual capability as a selection constraint
CV collections often contain mixed languages, code-switching, local degree names, international employer names, and non-English section headings. Gemma 4 advertises support for more than 140 languages [4]. Qwen3 advertises more than 100 languages and dialects [8] and is therefore the closest multilingual alternative in the table. Mistral Small supports dozens of languages [10], while Llama 4 officially supports 12 languages even though its pretraining includes a broader set [11]. Smaller English-centric models may perform well on English extraction but should not be assumed to preserve equivalent multilingual recall.
Multilingual evaluation must be empirical and field-specific. Overall language support does not guarantee accurate recognition of CV abbreviations, degree nomenclature, date formats, or transliterated names. The pooled reports do not provide a language-stratified breakdown, so the current analysis cannot quantify whether multilingual documents are a major source of omission. Future evaluation should record document language and report precision, recall, and completeness by language group.
9.3 Model substitution must be empirically validated
A valid model comparison must use the same PDF text, prompt, schema, reasoning mode, context, sampling parameters, constrained-decoding settings, and evaluation protocol. It should report both quality and speed. A smaller model may be preferable if it preserves precision and materially improves throughput, but a lower recall model would amplify the principal weakness already observed. Conversely, a reasoning-oriented model may improve recall but increase latency and unsupported inference.
Task-specific fine-tuning or low-rank adaptation may outperform a generic model swap if a representative training set can be created without leaking sensitive CV data. Any fine-tuning dataset should preserve multilingual diversity and include hard negatives such as skills appearing only in hobbies, dates detached from roles, and ambiguous education/certification boundaries.