llama.cpp auto-translate: add Qwen 3 4B Instruct + Qwen 3 8B#11033
Merged
Conversation
The curated translate-model list was four quants of the same model family (TranslateGemma). Add Qwen 3 as an alternative family — strongest open model for CJK languages, competitive elsewhere — so users hitting Gemma's quirks (refusals on adult-themed dialogue, formatting drift) have a real fallback rather than just a different quant of the same underlying model. Two new entries: - Qwen 3 4B Instruct (Q4_K_M, 2.5 GB) — dedicated instruct-only variant from the Qwen3-Instruct-2507 series. - Qwen 3 8B (Q4_K_M, 4.7 GB) — original hybrid Qwen3-8B. Qwen never released a separate Qwen3-8B-Instruct-2507; on the hybrid model, --no-jinja + --chat-template chatml bypasses the embedded Jinja template's enable_thinking logic so output is clean translation rather than <think>...</think> reasoning blocks. GGUF sources are bartowski's repos — most prolific, well-maintained quantizer. TranslateGemma 4B Q4_K_M stays as the default first-pick. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds Qwen 3 as an alternative model family in the curated llama.cpp auto-translate model list. Until now the list was four quantizations of the same model (TranslateGemma); users hitting Gemma's quirks (occasional refusals on adult-themed dialogue, formatting drift, weaker CJK quality) had no real fallback. Qwen 3 is the strongest open model for Japanese/Chinese/Korean translation in 2026 and is competitive on European pairs.
Two new entries in
LlamaCppServerManager.TranslateModels:bartowski/Qwen_Qwen3-4B-Instruct-2507-GGUFbartowski/Qwen_Qwen3-8B-GGUFWhy bartowski as the GGUF source
The two most reliable GGUF authors are ggml-org (official llama.cpp org) and bartowski. ggml-org's
Qwen3-4B-Instruct-2507-Q8_0-GGUFonly ships the 8 GB Q8_0 quant, and they don't ship Qwen3-8B-Instruct as separate GGUFs at all. bartowski ships every standard quant for both, is the highest-trafficked GGUF maintainer on HF, and tracks upstream releases promptly.Why Qwen3-8B (hybrid) and not Qwen3-8B-Instruct-2507
Qwen never released a separate
Qwen3-8B-Instruct-2507— only the 4B and 30B-A3B sizes got the Instruct/Thinking split. For the 8B size the original hybridQwen3-8Bis what's available.The existing wiring already handles this correctly: setting
ChatTemplate: "chatml", NoJinja: truetranslates to--no-jinja --chat-template chatmlon the llama-server command line. That bypasses the embedded Jinja template'senable_thinkinglogic and feeds the model a plain chatml prompt, so output is clean translation rather than<think>...</think>reasoning blocks.What stays the same
Test plan
<think>blocks, no system-token leakage🤖 Generated with Claude Code