Contents
In brief
Cloud models translate books and articles better every month, but budget, provider rules, and copyright make that channel unreliable. On Habr, the author ran local models on a machine with an RTX 5060 Ti 16 GB, a Core Ultra 7 265, and 96 GB of RAM. The goals were a fast in-browser page translation and an overnight book translation. Claude and OpenAI were the judges, and Claude also wrote the reference, on a 0–100 scale. Qwen3.8 Flash-Next led on quality and does not fit on the graphics card. For everyday speed on 16 GB, Gemma 4 26B in a quant that sits entirely on the GPU won.
What happened
Everything was translated into Russian from English, Spanish, Chinese, and Hebrew. The ordinary set included a WSJ column, a PubMed article, kernel docs, a thread with comments, contemporary prose, Lorca's poems, and a chapter of Amos Oz. A separate stress set covered Ben Sira, a religious passage, Kabbalah, old Chinese medical and Daoist texts, and a chapter of Finnegans Wake. Kabbalah was the hardest: the local models slumped on it.
The author kept few finalists. Qwen3.8 Flash-Next, a mixture of experts with about 125 billion weights and roughly 7 billion active, leads on quality: English 87.6, Spanish 92, Chinese 88, Hebrew 90, classical Hebrew 81, eight points ahead of the next model. It wants 80–95 GB of memory plus a GPU. The IQ4_XS quant ran near 15 tokens a second, the lighter IQ3_XXS near 21, and the score gap was noise. Literary text without a separate edit pass was only average. Reasoning mode left untranslated scraps in held-out classical Chinese.
Gemma 4 31B writes Russian better than the others and lifts a weak draft by 5–14 points, but on this machine it manages about 6–7 tokens a second because half the layers sit in RAM. Gemma 4 26B in Google's QAT Q4 quant fits entirely in 16 GB and runs near 113 tokens a second, scoring 80–88 on ordinary text and web pages. It falls apart on classics: classical Hebrew near 49, classical Chinese near 54. Index Translate 35B is stronger than the others on classical Chinese (about 80 in Q8) and works as a fast draft for the 31B edit pass. Smaller models and narrow translation variants were cut in early rounds.
Why it matters
Speed and quality are not one axis. A browser page cannot wait a minute. A book can run overnight. A 16 GB card is not a 16 GB model. The quality leader lives in system memory, and the GPU holds expert layers. With 32 GB of RAM that leader simply will not start, whatever the Spanish score of 92 says.
The judges are themselves cloud models, and so is the reference. This compares local translation with Claude, not with a human translator. The author says outright that style in difficult prose and local realia still slip, and that cloud models are still stronger on books. A local setup matters where the channel can close over budget, policy, or rights.
In practice
- For pages and ordinary text on this card, look at Gemma 4 26B in the quant that sits fully on the GPU. Reasoning adds about a point and a half and stretches time four or five times.
- For quality, if you have 80–95 GB of memory, keep Qwen3.8 Flash-Next. A Russian instruction gained about five points on classical texts. Reasoning did not help everywhere: it failed on held-out classical Chinese.
- Run literary text as a draft plus a separate Gemma 4 31B pass. The official English system text with a style instruction beat a Russian instruction and a "literary" sampler.
- Give classical Chinese to Index Translate, in short fragments. On Hebrew, a thousand-token fragment was translated only at the end. Pieces of 300–600 tokens fixed that.
- Do not expect a local model to replace the cloud for a book you plan to sell. The author keeps cloud models for that goal.
Takeaway
On a modest gaming card, local translation is already good enough to understand a page and a draft. The quality leader is a large model in RAM, not whatever fit in 16 GB. Classics and Kabbalah still break the finalists.
Comments