Quantised builds of the flagship conversational Welsh model in the EryriLabs Gemma 4 line, for llama.cpp and everything built on it.

The adapter, the evaluation record and the data cards live in the source repository EryriLabs/gemma-4-12b-cymraeg-v6. The bf16 merged weights are not published anywhere: they are regenerable from the base model plus the three adapters named there.

Method first: 100 fresh conversational Welsh prompts were authored for this test and machine checked for zero overlap against every earlier prompt set and against all training data of every model involved. Three models answered under their own documented settings. Scoring was blind (shuffled slots, deterministic seed) on a rubric fixed before any answer existed: helpfulness, correctness and naturalness of the Welsh, register, and fitness of length for a chat turn. The judge is a language model, not a native speaker; its per prompt scores, its uncertainty list, and the full prompt sets are published in eval/ so anyone can re mark the papers.

On that benchmark this model won 66 prompts to 33 with 1 tie against BangorAI/Mistral-7B-Cymraeg-Welsh-v2 (win rate 66.7 percent, 95 percent CI 56.9 to 75.2, p = 0.0012), and won a separate 14 turn dialogue test 4.34 to 3.74 with the strongest cross turn memory of any model tested. In the same set of experiments BangorAI's model won a 20 prompt mixed category exam 13 to 7 at each model's shipping settings and remains the stronger long form prose writer, with the most polished Welsh sentence by sentence (language quality 4.52 vs 4.18). The difference is behavioural: in the conversational hundred, BangorAI's replies included assistant style boilerplate 25 times, refused a reasonable request 17 times and missed the ti/chi register 15 times, against 0, 1 and 5 for this model.

BangorAI's model is the strongest other open Welsh model available and the only 7B class comparison; this work also trains on open data released by techiaith at Bangor University and by the Welsh public sector. Diolch o galon i bawb.

Lineage: google/gemma-4-12b-it, continued pretraining on about 307 million tokens of open Welsh text (the full MADLAD-400 Welsh split plus the techiaith corpora; embeddings trained in the first stage), then conversational supervised fine tuning. See DATA-CARD files and cpt_progress.json for exact coverage.

Q6_K only. It is the quant this release is validated around, and the one to use unless there is a specific reason not to. Q8_0 and Q4_K_M are not published here, purely because the run that produced this release had no budget left to convert and gate them properly; both are regenerable from the source repository at any time, and shipping an ungated 4 bit build would put the quantisation defect described below in front of users under this model's name.

Every quant published here passed a corruption gate before upload. Sizes, sha256 hashes and the per quant corruption counts are in MANIFEST.json.

Thinking off. Temperature 0.3, top_p 0.9, min_p 0.05, repetition penalty 1.1. A Welsh system prompt noticeably improves Welsh only behaviour in long conversations, and the multi turn evaluation in the source repository measures the difference with and without one.

Blind LLM as judge (Opus), identical prompts and settings for every arm. The judge is not a native speaker: structural findings are reliable, fine style judgements less so. Full report and per response scores are in eval/.

Welsh, 40 prompts:

Head to head with BangorAI/Mistral-7B-Cymraeg-Welsh-v2, the strongest other open Welsh model and the only 7B class comparison (diolch to BangorAI and to techiaith at Bangor University, whose open data trained much of this model): single turn pairwise 15 wins, 5 ties, 20 losses (BangorAI keeps the single turn edge; the gap has narrowed from 14-21 for v2 to 16-18 for v6 greedy). Multi turn conversation, 14 fixed turns with history: v6 4.34 vs BangorAI 3.74, per turn 8-2-4, with cross turn memory the largest gain (3.92 to 4.46). English is fully retained (4.90 fluency, 20/20 good English).

v6 vs v5 pairwise: 16 wins, 16 ties, 8 losses. The additional immersion bought coherence and fact repair more than sentence elegance.

Known limitations: long form answers past ~150 tokens can leak English or loop under greedy decoding (cap generation at ~160 tokens and use the sampling below, which removes every observed catastrophic failure); rare unprovoked refusals (2 of 80 responses); it performs mutations far better than it explains them; numeric facts should be grounded via the system prompt.

Recommended settings: temperature 0.3, top_p 0.9, min_p 0.05, repetition_penalty 1.1, max_new_tokens 160, thinking off.

Apache 2.0, matching the base model google/gemma-4-12b-it.