Cohere's 218B MoE Scores 83.6 on WMT26, Beats DeepL and Google Translate

Cohere and Cohere Labs have released North Small Translate, an open-weight sparse Mixture-of-Experts model built specifically for machine translation across 50 languages, from Albanian to Vietnamese. The model carries 218B total parameters but activates only 25B per token — roughly 11.5% of weights — keeping per-token compute aligned with a 25B dense model while memory requirements still track the full 218B.

The release matters because translation has historically been handled by either proprietary APIs with opaque quality floors or general-purpose LLMs not post-trained for the task. North Small Translate sits in a different category: purpose-built, open-weight, and available via three distinct licensing paths — free on Cohere's API up to rate limits, self-hosted under CC BY-NC 4.0 for non-commercial use, or commercially licensed through Cohere Model Vault. The model was developed in partnership with RWS, whose Language Weaver team contributed language-science expertise alongside Cohere's training infrastructure.

Architecture: MoE routing meets hybrid attention

The decoder-only transformer uses 128 experts per MoE layer, with a sigmoid router scoring all 128 and retaining the top 8 per token. Those 8 expert weights are then normalized over the selected top-k set. Separately, shared experts run on every token regardless of routing, adding consistent representational capacity across all inputs.

The attention stack interleaves sliding-window and global layers in a 3:1 ratio. Sliding-window layers use RoPE with a 4,096-token window; every fourth layer is global, attends to the full context, and applies no positional embeddings. Cohere notes this attention layout originated in Command A. Both input and output context windows cap at 16K tokens — a limit that deserves scrutiny for book-length document translation, where chunking reintroduces the coherence problems the 16K window was meant to solve.

Benchmarks: strong numbers with an important caveat

Cohere's scoring bands classify 80–100 as "perfect or minor errors only." Both the standard model at 83.60 and the agentic multi-pass variant at 84.36 fall in that range. Regional breakdowns show the standard model scoring 82.17 on EU languages against Gemma 4 31B's 72.73 — a 9.44-point gap. South Asia narrows considerably: 86.16 for North versus 88.04 for Gemma, where Gemma holds a slight lead.

The caveat is structural: all scores are vendor-reported and judged by GPT-5.6-Sol rather than a neutral third-party evaluation. Independent WMT26 results don't yet exist for this model. Teams should treat these figures as directional until external replication appears.

Speed, cost, and hardware footprint

At low concurrency on identical hardware, Cohere's internal tests show 112 output tokens per second for North Small Translate against 81 for Gemma 4 31B. At high concurrency, those figures fall to 39 and 30 respectively — Cohere characterizes this as up to 1.4x higher throughput. Long-document quality, measured per paragraph with xCOMET-XL across 2 book chapters in a single call, shows a sharper separation: North scores 48.9 against Google Translate's 21.3 and Gemma's 19.4.

On cost, Cohere's figures show the model at $0.000676 per task averaging 661 tokens, versus $0.038928 for Gemini 3.1 Pro Preview (high) — approximately 58x more expensive — and $0.004525 for Qwen 3.5 397B A17B. Self-hosting requires, at minimum, 1x B200 or 2x H100 for the NVFP4 W4A16 checkpoint, scaling to 4x B200 or 8x H100 for BF16. The three checkpoints Cohere publishes are the same ones it serves in production.

AI Mastery analysis

The 218B / 25B active parameter split is the practical crux for anyone evaluating this model for production pipelines. Per-token FLOPs scale with the 25B active count, which is favorable for throughput and per-request cost. But GPU memory allocation must cover all 218B, which makes the hardware minimums above non-negotiable. A team running 4x H100s in FP8 mode is committing substantial infrastructure for a single-purpose translation endpoint — a very different calculation than routing traffic to an API. This is exactly the pipeline architecture tradeoff that determines whether a specialized model delivers its efficiency gains or shifts cost to infrastructure.

The agentic variant — which runs a multi-pass error-detection and correction loop — adds 0.76 WMT26 points over the single-pass model (84.36 vs. 83.60), but introduces latency and additional token cost that Cohere doesn't fully quantify. For high-volume batch translation, that overhead may not be justified. For legal, medical, or regulatory documents where error correction has compounding value, the agentic mode is the more defensible choice.

The RWS collaboration is notable beyond marketing. Language Weaver has operated commercial translation pipelines for decades. That domain expertise, embedded during post-training rather than added at inference time, likely explains the long-document xCOMET-XL scores more than the raw architecture does. North Small Translate signals that specialized post-training for a narrow task — even at MoE scale — can outperform much larger general-purpose models on that task's specific metrics. The open weights and concrete hardware minimums give engineering teams the information needed to test that claim independently.

Primary source

Cohere Releases North Small Translate: A 218B MoE Translation Model That Scores 83.6 on WMT26 Across 50 Languages — MarkTechPost