Local AI by VRAM Tier - 8GB | 12GB | 16GB | 24GB | 32GB
Deep dives: Chat | Coding | Translation | Vision | Speech | Agents
Sending a document to Google Translate or a hosted translation API means sending the text itself to someone else’s servers, and having it metered on the way through. A translation model running on local hardware keeps the text on the machine, works offline, and does not phone home.
Updated August 2026. Refreshed monthly as models and runtimes change.
Local translation splits into two camps: dedicated translation models built for the task, and general LLMs that translate well as a side effect. Google’s TranslateGemma, released in January 2026, is still the strongest dedicated family available for download, and it is still built on the Gemma 3 architecture rather than the newer Gemma 4. What follows is the picture at each VRAM tier, with the licence attached to every option, because licence terms disqualify more deployments than VRAM does.
Corrections in This Update
Three things on this page were wrong or a generation out of date. All three are fixed below.
LibreTranslate is AGPL-3.0, not MIT. An earlier version of this page carried a single MIT licence covering the Argos Translate and LibreTranslate pair. That conflated two projects with two different licences, and the difference matters. Argos Translate, the translation library, is genuinely permissive: it is dual-licensed MIT and CC0. The LibreTranslate server that wraps it is published under the GNU AGPL-3.0. AGPL is a network copyleft, which is to say it reaches users who never receive a binary and only ever see an HTTP response. Running a modified LibreTranslate as a service others can reach over a network carries source-availability obligations that embedding Argos as a library does not. If the plan is a hosted translation endpoint, that single distinction decides the design.
Qwen-MT is a hosted API product, not a downloadable model. Qwen-MT is an Alibaba Cloud service reached over their API. No open weights for Qwen-MT have been published, so nothing in the Qwen-MT announcement can be installed, quantised, or run on the hardware described on this page, and none of its results are evidence about a local model. An earlier version of this page cited it as support for a local Qwen recommendation. That was misleading. Every Qwen recommendation below rests only on the open-weight Qwen3.5 and Qwen3.6 checkpoints, which are separate Apache 2.0 releases.
Qwen 3 references now point at Qwen3.5 and Qwen3.6. The original Qwen 3 generation has been superseded twice. Model names, quantisations and file sizes below are for the current generation.
Licence Before Benchmark
For translation specifically, the licence column decides more outcomes than the benchmark column. The widest-coverage dedicated models are the ones least likely to be shippable.
| Project | What it is | Licence | Practical effect |
|---|---|---|---|
| TranslateGemma 4B / 12B / 27B | Dedicated translation, Gemma 3 based | Google’s Gemma Terms of Use plus a separate Prohibited Use Policy | Not a blanket open licence. Custom terms with explicit use restrictions that travel downstream |
| Argos Translate | CPU translation library | MIT and CC0, dual | Permissive. No obligation triggered by serving it |
| LibreTranslate | Self-hosted API and web UI wrapping Argos | AGPL-3.0 | Network copyleft. Serving a modified version triggers source obligations |
| NLLB-200 1.3B / 3.3B | Dedicated translation, 200 languages | CC-BY-NC-4.0 | Non-commercial only |
| Aya Expanse 8B / 32B | Multilingual general model | CC-BY-NC-4.0 | Non-commercial only |
| Qwen3.5 and Qwen3.6 | General LLMs | Apache 2.0 | Commercial use permitted |
| MADLAD-400 3B | Dedicated translation, 400+ languages | Apache 2.0, per the model card | Commercial use permitted, ungated |
NLLB-200 and Aya Expanse are both CC-BY-NC-4.0. Between them they cover 200 languages and the strongest low-resource quality in this comparison, and neither can be used commercially. That is the most common unpleasant surprise in this category.
Two Approaches to Local Translation
Dedicated Translation Models
Trained for translation, without the overhead of general chat capability.
| Model | Params | Languages | Size | Licence |
|---|---|---|---|---|
| TranslateGemma 4B | 4B | 55 | 3.3 GB (Ollama default tag) | Gemma Terms + Prohibited Use Policy |
| TranslateGemma 12B | 12B | 55 | 8.1 GB (Ollama default tag) | Gemma Terms + Prohibited Use Policy |
| TranslateGemma 27B | 27B | 55 | 17 GB (Ollama default tag) | Gemma Terms + Prohibited Use Policy |
| NLLB-200 1.3B | 1.3B | 200 | ~2 GB | CC-BY-NC-4.0 |
| NLLB-200 3.3B | 3.3B | 200 | ~4 GB | CC-BY-NC-4.0 |
| MADLAD-400 3B | 3B | 400+ | ~4 GB | Apache 2.0 |
| Argos Translate | Small | 60+ | CPU only | MIT / CC0 |
General LLMs Used for Translation
General-purpose models that produce more natural, context-aware output, which often matters more than literal accuracy for prose.
| Model | Params | Languages | Size | Licence |
|---|---|---|---|---|
| Qwen3.5-4B | 4B | 201 (vendor claim) | 2.74 GB at Q4_K_M | Apache 2.0 |
| Qwen3.5-9B | 9B | 201 (vendor claim) | 5.68 GB at Q4_K_M, 9.53 GB at Q8_0 | Apache 2.0 |
| Qwen3.5-27B | 27B | 201 (vendor claim) | 16.74 GB at Q4_K_M, 22.45 GB at Q6_K | Apache 2.0 |
| Qwen3.6-27B | 27B | 201 (vendor claim) | 17 GB (Ollama q4_K_M tag) | Apache 2.0 |
| Aya Expanse 8B | 8B | 23 | ~6 GB | CC-BY-NC-4.0 |
| Aya Expanse 32B | 32B | 23 | ~22 GB | CC-BY-NC-4.0 |
A note on the Qwen numbers: those are static GGUF weight sizes from the Hugging Face Hub at the stated quantisation, and they exclude the KV cache. Real usage climbs with context length, so anything sitting within a gigabyte or two of a tier boundary should be read as “fits with minimal context room” rather than “fits comfortably”. The 201-language figure is Qwen’s own claim, carried forward from the previous generation.
TranslateGemma: Still the Default, With a Licence Caveat
TranslateGemma arrived in January 2026, fine-tuned from Gemma 3 for translation across 55 languages spanning high-, mid- and low-resource pairs. Gemma 4 has since shipped as a general model family, but TranslateGemma has not been rebased onto it; the translation models are still Gemma-3-based.
All three sizes are in Ollama’s official library:
ollama pull translategemma:4b # 3.3 GB
ollama pull translategemma:12b # 8.1 GB
ollama pull translategemma:27b # 17 GBThose are the sizes Ollama publishes for the default tags, which is a firmer number than a community estimate and matches the figures this page has carried since March.
The licence is not Apache or MIT, and the weights are gated. TranslateGemma inherits Google’s custom Gemma Terms of Use together with a separate Prohibited Use Policy, and all three Hugging Face repositories require manual access approval before you can download anything. The Ollama tags are not gated. The weights are downloadable and commercial use is contemplated, but the terms impose use restrictions that a permissive open-source licence does not, and Google requires those restrictions to be passed on with any distribution or derivative. Anyone building a product on TranslateGemma should read both documents rather than assume “open weights” means “open licence”.
Quality: What Google’s Own Report Claims
Every WMT24++ figure below comes from the TranslateGemma technical report, published by Google alongside the models. These are vendor self-reported results. No independent reproduction of them was located for this update.
| Model | MetricX | COMET22 | Notes |
|---|---|---|---|
| TranslateGemma 27B | 3.09 | 84.4 | Best of the dedicated line |
| TranslateGemma 12B | 3.60 | 83.5 | Ahead of the Gemma 3 27B baseline in the same report |
| Gemma 3 27B (baseline) | 4.04 | 83.1 | General-model baseline |
| TranslateGemma 4B | 5.32 | 80.1 | Strong for the size |
| Gemma 3 12B (baseline) | 4.86 | 81.6 | General-model baseline |
| Gemma 3 4B (baseline) | 6.97 | 77.2 | Functional but limited |
MetricX is lower-is-better, COMET22 higher-is-better. The only comparisons the report makes are against its own Gemma 3 baselines, so this table says nothing about how the models compare with hosted commercial translation services. No independent benchmark placing them against those services was found in this pass.
By VRAM Tier
8GB VRAM {#8gb}
GPUs: RTX 4060, RTX 3060 8GB, RTX 3070
| Model | Size | Languages | Licence |
|---|---|---|---|
| TranslateGemma 4B | 3.3 GB | 55 | Gemma Terms + Prohibited Use Policy |
| Qwen3.5-4B | 2.74 GB at Q4_K_M | 201 (claim) | Apache 2.0 |
| NLLB-200 1.3B | ~2 GB | 200 | CC-BY-NC-4.0 |
| Argos Translate | CPU only | 60+ | MIT / CC0 |
Best pick: TranslateGemma 4B. At 3.3GB it leaves roughly 4.7GB free for context and clears the Gemma 3 4B baseline by a wide margin in Google’s own reporting (80.1 against 77.2 COMET22). Purpose-built translation still beats a general model of the same size. Check the Gemma terms before putting it in a product.
Permissively licensed alternative: Qwen3.5-4B. Apache 2.0, 2.74GB at Q4_K_M, and no use policy to read. Output reads more like prose and less like a gloss, at the cost of the translation-specific tuning.
Maximum language coverage: NLLB-200 1.3B. 200 languages in around 2GB. The translations are more literal than TranslateGemma’s. Useful for rare pairs where TranslateGemma has no coverage at all, and non-commercial only under CC-BY-NC-4.0.
Zero-GPU option: Argos Translate. Runs on CPU with no GPU at all, 60+ language pairs, MIT and CC0 dual-licensed as a library. Quality sits a clear step below the LLM options, but zero VRAM means it runs alongside anything else. If it gets wrapped in LibreTranslate and exposed as a service, the AGPL-3.0 terms described above apply to the wrapper.
For all use cases at this level, see the 8GB VRAM complete guide.
12GB VRAM {#12gb}
GPUs: RTX 3060 12GB, RTX 4070
| Model | Size | Languages | Licence |
|---|---|---|---|
| TranslateGemma 12B | 8.1 GB | 55 | Gemma Terms + Prohibited Use Policy |
| Qwen3.5-9B | 5.68 GB at Q4_K_M | 201 (claim) | Apache 2.0 |
| Aya Expanse 8B | ~6 GB | 23 | CC-BY-NC-4.0 |
| NLLB-200 3.3B | ~4 GB | 200 | CC-BY-NC-4.0 |
Best pick: TranslateGemma 12B. This is the inflection point. At 8.1GB it reports MetricX 3.60, ahead of the Gemma 3 27B baseline in the same technical report while using far less VRAM. For the 55 supported languages this is the strongest quality-per-gigabyte on the page.
CJK and general prose: Qwen3.5-9B, 5.68GB at Q4_K_M, Apache 2.0. Previous versions of this section pointed at Qwen-MT results to justify a Qwen pick here. That was an error: Qwen-MT is an Alibaba Cloud API product with no published open weights, and its scores say nothing about a downloadable checkpoint. The case for Qwen3.5-9B is the open-weight model on its own terms, available through Ollama under Apache 2.0 with the broadest claimed language list of anything here.
Running both at once does not fit this tier: TranslateGemma 12B plus Qwen3.5-9B is about 13.8GB of weights before any KV cache, over a 12GB card. The combination that does fit is TranslateGemma 4B plus Qwen3.5-9B at roughly 8.98GB, leaving room for context.
Low-resource languages: Aya Expanse 8B. Cohere’s model covers 23 languages with a focus on underserved ones. The Aya Expanse paper reports win rates of up to 76.6% against Gemma 2, Qwen 2.5 and Llama 3.1 on Arena-Hard-Auto translated into 23 languages, which is Cohere’s own reporting rather than an independent result. Worth testing for Arabic, Hindi or Indonesian, and CC-BY-NC-4.0, so non-commercial only.
For all use cases at this level, see the 12GB VRAM complete guide.
16GB VRAM {#16gb}
GPUs: RTX 4060 Ti 16GB, RTX 5060, Arc A770
| Model | Size | Languages | Licence |
|---|---|---|---|
| TranslateGemma 12B | 8.1 GB | 55 | Gemma Terms + Prohibited Use Policy |
| Qwen3.5-9B at Q8_0 | 9.53 GB | 201 (claim) | Apache 2.0 |
| TranslateGemma 4B + Qwen3.5-9B at Q4_K_M | 3.3 + 5.68 GB | 55 and 201 | Mixed |
Best pick: TranslateGemma 12B with headroom. Same model as the 12GB tier, now with close to 8GB spare. That headroom buys context rather than quality: whole documents can go through in one pass, which is what keeps terminology consistent across a long text.
Higher-precision general LLM: Qwen3.5-9B at Q8_0, 9.53GB. For literary work, marketing copy or anything where tone carries as much as meaning, a general model at near-lossless quantisation reads better than a dedicated translator. It is slower and less consistent on technical and formal text.
Dual setup: TranslateGemma 4B for quick, literal passes plus Qwen3.5-9B at Q4_K_M for context-heavy work, 8.98GB of weights combined, with the rest of the card left for context.
For all use cases at this level, see the 16GB VRAM complete guide.
24GB VRAM {#24gb}
GPUs: RTX 3090, RTX 4090
| Model | Size | Languages | Licence |
|---|---|---|---|
| TranslateGemma 27B | 17 GB | 55 | Gemma Terms + Prohibited Use Policy |
| Qwen3.5-27B | 16.74 GB at Q4_K_M | 201 (claim) | Apache 2.0 |
| Aya Expanse 32B | ~22 GB | 23 | CC-BY-NC-4.0 |
| Qwen3.6-27B | 17 GB (Ollama q4_K_M tag) | 201 (claim) | Apache 2.0 |
Best pick: TranslateGemma 27B. MetricX 3.09 and COMET22 84.4 in Google’s own report, the best dedicated-translation figures available, at 17GB from Ollama’s default tag. That leaves around 7GB for context or a second model.
Best commercially usable pick: Qwen3.5-27B at Q4_K_M, 16.74GB, Apache 2.0. Fits the same tier with no use policy attached, which is the deciding factor for anything shipping to customers.
Newest generation: Qwen3.6. The Qwen3.6 line published in April 2026 exists only at 27B and 35B-A3B, both Apache 2.0. There is no 4B or 9B counterpart, so it cannot fill the lower tiers, and no translation-specific benchmark for it was located for this update. Treat it as the newest option at 24GB and above rather than a proven translation upgrade. Ollama now lists its dense 27B at 17GB, the same as Qwen3.5-27B, so it drops into this tier without changing the arithmetic.
Multilingual specialist: Aya Expanse 32B at around 22GB. Cohere’s paper reports that the 32B model beats Llama 3.1 70B, twice its size, with a 54.0% win rate across those 23 languages, again self-reported. A chrF++ figure quoted in an earlier version of this page could not be traced to the paper or the model card and has been removed. For Arabic, Hindi, Persian, Indonesian and similar languages this is the strongest option on the page, and CC-BY-NC-4.0 keeps it out of commercial work.
For all use cases at this level, see the 24GB VRAM complete guide.
32GB VRAM {#32gb}
GPUs: RTX 5090
| Configuration | Size | Notes |
|---|---|---|
| TranslateGemma 27B + Qwen3.5-9B at Q4_K_M | 17 + 5.68 GB | Dedicated translator plus a general model |
| Qwen3.5-27B at Q6_K | 22.45 GB | Single model, near-lossless, Apache 2.0 |
Best pick: TranslateGemma 27B alongside Qwen3.5-9B. About 22.7GB of weights combined, leaving real context headroom on a 32GB card. Route literal and technical pairs to the dedicated model and prose to the general one.
Ollama publishes higher-precision TranslateGemma 27B builds alongside the 17GB default: a q8_0 tag at 30GB and a bf16 tag at 55GB. The q8_0 build fits a 32GB card on its own, with no room for a second model. Earlier versions of this page listed a Q6_K TranslateGemma figure that no primary source supports; it has been removed.
Single-model maximum: Qwen3.5-27B at Q6_K, 22.45GB. Near-lossless, Apache 2.0, and the widest claimed language list of anything that fits one card.
For all use cases at this level, see the 32GB VRAM complete guide.
Cross-Tier Summary
| Tier | Best pick | Licence of that pick | Commercially unrestricted alternative |
|---|---|---|---|
| 8GB | TranslateGemma 4B | Gemma Terms + Prohibited Use Policy | Qwen3.5-4B, Apache 2.0 |
| 12GB | TranslateGemma 12B | Gemma Terms + Prohibited Use Policy | Qwen3.5-9B, Apache 2.0 |
| 16GB | TranslateGemma 12B with headroom | Gemma Terms + Prohibited Use Policy | Qwen3.5-9B at Q8_0 |
| 24GB | TranslateGemma 27B | Gemma Terms + Prohibited Use Policy | Qwen3.5-27B at Q4_K_M |
| 32GB | TranslateGemma 27B + Qwen3.5-9B | Mixed | Qwen3.5-27B at Q6_K |
When to Use What
TranslateGemma for straightforward translation across its 55 languages. Best accuracy per gigabyte, fastest processing, and custom Google terms with a prohibited-use policy attached.
Qwen3.5 and Qwen3.6 for prose, literary and creative translation, the broadest claimed language list, and anything that has to be commercially unrestricted. Apache 2.0.
Aya Expanse for low-resource languages and pairs where mainstream models degrade. Non-commercial only.
NLLB-200 for maximum language coverage, especially rare pairs. Literal but accurate, and non-commercial only.
Argos Translate for zero-GPU and embedded use, under MIT and CC0.
LibreTranslate for a self-hosted endpoint with a web UI and API, provided AGPL-3.0 is acceptable for the deployment. If the service is modified and exposed to others over a network, plan for the source obligations.
On the Watch List, Not Ready to Recommend
Tiny Aya (Cohere, 3.35B, 70+ languages) is a small general multilingual base model, not a translation-tuned one, so it is not a drop-in replacement for anything above. Its licensing is also inconsistent between its own sources: the Hugging Face model card lists cc-by-nc, while the accompanying arXiv paper states CC-BY-NC-ND-4.0. Those are not the same terms, since the ND variant additionally restricts distributing derivatives. Until the card and the paper agree, no single licence can honestly be stated for it.
Meta Omnilingual MT (arXiv:2603.16309) describes translation across 1,600 languages and is, at the time of this update, a research paper. Downloadable open weights for it could not be confirmed in this pass, so it is not something to plan a local pipeline around yet.
Quick Start
# Dedicated translation
ollama pull translategemma:12b
ollama run translategemma:12b "Translate to French: The local AI revolution is making translation private."
# General LLM alternative, Apache 2.0
ollama pull qwen3.5:9b
# Batch translation over the local API
curl http://localhost:11434/api/generate -d '{
"model": "translategemma:12b",
"prompt": "Translate to German: Privacy matters more than convenience.",
"stream": false
}'Model tags change as the libraries update, so check the Ollama listing for the current tag rather than trusting a pinned name. For a full LibreTranslate setup with web interface and API, see the Self-Host LibreTranslate guide, and read the AGPL-3.0 terms above before exposing that endpoint to anyone else. For an LLM-backed drop-in that speaks the same API, LTEngine is the project’s own route.
Honest Limits
Local translation still has real gaps against hosted commercial services:
- Rare language pairs. Quality drops sharply on uncommon combinations. Even NLLB-200’s 200 languages are thin for many of them
- Idioms and cultural context. Dedicated models translate too literally. General LLMs do better and are still not reliable
- Consistency across long documents. Keeping terminology stable through a 50-page translation needs manual review regardless of model
- Throughput. Bulk translation is bounded by one GPU rather than a provider’s fleet
- Licensing. The two models with the widest language coverage, NLLB-200 and Aya Expanse, are both non-commercial, and the strongest dedicated family carries custom Google terms rather than an open-source licence
For the 55 languages TranslateGemma covers, the quality gap keeps narrowing on Google’s own numbers. For everything else it depends heavily on the specific pair, and on whether the licence lets the result leave the machine.
All eleven local-AI guides, by hardware tier and by task, are indexed at Local AI by VRAM: Which Models Fit Your GPU.