Tessera-27B-B-CJK, Zellige AIs new finetune of Qwen with novel approaches to how we think about token savings
Tessera-27B-B-CJK is a Qwen3.8-27B coding agent that thinks in compressed CJK, and it spends fewer tokens doing it.
Proof1 community thread
Key points
- Tessera is our lineup of models and tunes that try new ways to save tokens, run more efficiently and show other parts of our research.
- So we taught Qwen3.8-27B to think less, and to think in compressed CJK, a shorthand that mixes English, Chinese words and symbols.
- Here is one real turn from a held-out task, 27 tokens before two tool calls:
- 0.50-0.64x the base model's thinking tokens per agent turn (typical turn 0.496x, 90% CI 0.385-0.624; token-weighted 0.644x)
Sources (1)
- [1]Tessera-27B-B-CJK, Zellige AIs new finetune of Qwen with novel approaches to how we think about token savingsr/LocalLLaMA (top, daily) · Oct 14, 05:35 PM
Tessera-27B-B-CJK is a Qwen3.8-27B coding agent that thinks in compressed CJK, and it spends fewer tokens doing it.
Tessera is our lineup of models and tunes that try new ways to save tokens, run more efficiently and show other parts of our research.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 14, 2026New model from Telperion: Qwen3.8-27B-MLA & Qwen3.8-27B-GLA-g2. 4x smaller KV cache, smarter, and more efficient reasoning.
- Oct 14, 2026NVIDIA/TensorRT-LLM v1.4.0rc0
- Oct 14, 2026mistralai/Timestral-209M
- Oct 13, 2026Run fast, private agents on-device with LFM2.5-24B-A2B
- Oct 13, 2026black-forest-labs/flux-3-image-klein-comfy
- Oct 13, 2026SonicBoom: 52 tok/s decode on a laptop RTX 3050 (llama.cpp: 45) with the same quantized kernels, the difference is a graph compiler. No Python.
