CosineAI/lumen-outpost-2026-04-27

Lumen Outpost

Lumen Outpost is a fine-tuned variant of Kimi-K2.6 produced by Cosine AI. It is trained on Cosine's proprietary dataset using custom-built graders designed to improve output quality across several dimensions:

• Stronger capability on niche and low-resource languages. Fine-tuning on targeted multilingual data improves fluency and correctness in languages that are underserved by the base model's pretraining distribution.
• Reduced slop. Trained against code quality graders that penalize dead code, duplication, unnecessary abstractions, and noisy comments. The model produces cleaner patches with less residual noise.
• Better conversational feel. Trained against conversational quality graders that reward concise and substantive updates, professional tone, and alignment between what the model says and what it does.

The base model is Moonshot AI's Kimi-K2.6, a 1T-parameter native multimodal MoE model with 32B active parameters per token. For full details on the base architecture, capabilities, and benchmarks, see the Kimi-K2.6 model card.

Model Details

Field	Value
Base model	Kimi-K2.6
Architecture	Mixture-of-Experts (MoE) — same architecture family as upstream Kimi-K2.6
Total parameters	1T
Active parameters	32B per token
Layers	61 (including 1 dense layer)
Experts	384 routed + 1 shared, 8 selected per token
Context length	256K tokens
Weight format	BF16 + INT4 packed MoE experts
Size on disk	~595 GB
Tokenizer	TikToken-based, 163,840 vocab
Vision encoder	MoonViT (400M params)

BF16 is used for attention and shared MLP weights. Routed MoE experts are stored as packed INT4 tensors. This checkpoint merges the lumen-outpost-2026-04-27 LoRA into Kimi-K2.6, including re-AWQ'd routed expert INT4 LoRA deltas packed back into Kimi-compatible INT4 tensors.

Serving

Use multi-GPU tensor parallelism.

vllm serve CosineAI/lumen-outpost-2026-04-27 \
  --served-model-name lumen-outpost \
  --api-key "$VLLM_API_KEY" \
  --trust-remote-code \
  --tensor-parallel-size 8 \
  --mm-encoder-tp-mode data \
  --enable-auto-tool-choice \
  --tool-call-parser kimi_k2 \
  --reasoning-parser kimi_k2 \
  --gpu-memory-utilization 0.95 \
  --dtype bfloat16

For additional deployment options (SGLang, KTransformers), refer to the base model deployment guide.

The chat template is included (chat_template.jinja) and supports both thinking and instant modes — same usage as base model. See the base model README for API usage examples including chat completion, vision input, tool calling, thinking mode toggles, and preserve-thinking mode.

Requirements

Software:

• Python >= 3.10
• transformers >= 4.57.1, <5.0.0 (same requirement as base model)
• tiktoken — required by the custom tokenizer (tokenization_kimi.py)
• tokenizers — required by tiktoken tokenizer internals
• numpy, Pillow, pydantic — required by vision processing code
• flash-attn >= 2.1 — optional but strongly recommended for attention performance. Without it, the model falls back to eager attention (functional but slow).
• mecord — optional, only needed for video input processing. Image-only and text-only usage does not require it.
• vllm, sglang, or ktransformers — for serving. Direct transformers generation is possible but not practical at this model size.

Hardware (minimum for inference):

• See the Kimi-K2.6 deployment guide

License

Please refer to the LICENSE.md file in this repository.

Acknowledgements

Lumen Outpost is built on Moonshot AI's Kimi-K2.6 base model. Credit to Moonshot AI for the Kimi-K2.6 architecture, training, and release. See the Kimi-K2.6 model card and technical blog for details. The underlying DeepSeek-V3 architecture is credited to DeepSeek.