Add CohereLabs/North-Mini-Code-1.0 recipe

[abtraore]

Adds a recipe for CohereLabs/North-Mini-Code-1.0 — Cohere's first open-weights agentic coding model (June 2026, Apache 2.0): 30B-total / 3B-activated MoE (128 experts, top-8 sigmoid routing, interleaved SWA-4096 + global attention), 256K context with 64K output. First CohereLabs recipe in the catalog, so this also registers the provider.

What's in the recipe

• Variants: BF16 default (TP=2, per Cohere's reference command) and the FP8 checkpoint (single-GPU, --moe-backend triton per the card, VRAM sized from the real 32 GB footprint since attention/router/lm_head stay high-precision).
• min_vllm_version 0.24.0, deliberately above the v0.21.0 arch registration: loading this checkpoint fails on v0.22.1/v0.23.0 with KeyError: 'layers.0.mlp.down_proj.weight' ([Bug]: cohere2_moe failing with vLLM>=v0.22.1 vllm#46366, a Transformers >= 5.10 config interaction fixed by [Cohere] Fix Cohere2MoE weight loading when using Transformers ≥5.10 vllm#44747, first in a stable release at v0.24.0). The guide documents the trap.
• cohere_melody>=0.9.0 dependency — the cohere_command4 tool/reasoning parsers import it at startup (the floor matches vLLM's own test requirements); the guide notes the official Docker images don't bundle it and shows the 2-line image extension.
• Card-recommended sampling (temp 1.0 / top_p 0.95), --max-model-len 320000 (the validated 256K-input + 64K-output window; config ceiling is 500000), and the card's benchmark headline numbers.
• meta.hardware: rtx_5090_2x: verified.

Tested

Verified end-to-end on 2x RTX 5090 (TP=2) with the recipe's exact FP8 command: boots with the full --max-model-len 320000 window (15.1 GiB weights/GPU, 12.1 GiB KV, 919,703-token pool), 206 tok/s single-stream, reasoning extracted into reasoning_content, tool-call round-trip clean. We also serve this FP8 checkpoint in production on 4x RTX 5090 (TP=4 + --enable-expert-parallel, 256K context).

node scripts/build-recipes-api.mjs passes (154 models); the synthesized default command matches Cohere's reference serve command exactly.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview	Aug 5, 2026 2:33am

[gemini-code-assist]

Code Review

This pull request introduces the model configuration for Cohere's North-Mini-Code-1.0, a 30B Mixture-of-Experts model, detailing its architecture, dependencies, features, and deployment variants. Additionally, Cohere Labs is registered as a provider in src/lib/providers.js. I have no feedback to provide on these changes.