The 7B-parameter text language model that serves as the reasoning/knowledge backbone of Moshi. Trained from scratch by Kyutai on 2.1T tokens of filtered English.

Architecture & training

  • Autoregressive Transformer: RMSNorm, RoPE positional embeddings, 4,096 context, FlashAttention, Gated Linear Units (SiLU gating).
  • 32k SentencePiece unigram tokenizer (digits split to single tokens, byte-backoff).
  • AdamW, fixed LR then cosine decay.
  • Data: 12.5% curated (Wikipedia, Wikibooks, StackExchange, peS2o scientific articles) + 87.5% CommonCrawl, heavily filtered (line-level dedup via bloom filter + fastText fuzzy dedup, fastText language ID, fastText 9-category quality classifier).

Results

On-par with or better than similarly-compute-budgeted ~7B models (MPT, Falcon, Llama 2, OLMo) across ARC, OBQA, HellaSwag, WinoGrande, PIQA, TriviaQA, NQ, MMLU; competitive even with Mistral/Gemma (≈3× more training compute) on some. MMLU 54.3 — drops to 49.7 once adapted into Moshi (the cost of audio training).

Table 2 — text LM evaluation (ARCe/ARCc, OBQA, HellaSwag, WinoGrande, PIQA, SIQA, TriviaQA Unfiltered/Wiki, NQ, MMLU; bold = best):

ModelARCeARCcOBQAHSWGPIQASIQATQANQMMLU
Helium79.655.953.676.370.079.451.059.9/72.623.354.3
MPT70.546.551.477.669.980.648.5–/61.220.830.8
Falcon73.747.553.076.368.980.347.2–/64.621.028.0
Llama 275.245.958.677.269.278.848.3–/72.125.745.3
OLMo67.242.550.075.569.877.5–/–52.0
Mistral80.554.952.281.074.282.247.062.5/–23.262.5
Gemma 181.553.252.881.272.381.251.863.4/–23.064.3

(Mistral & Gemma use ≈3× more training compute than Helium’s 2.1T tokens.)

In Moshi, Helium initializes the Temporal Transformer of the RQ-Transformer; 50% text-only batches are retained during Moshi pre-training to preserve its knowledge.

Source: Moshi (Kyutai paper).