A 9B coding model fine-tuned to reason internally in a compressed CJK register while emitting ordinary code and well-formed tool calls. On the full LiveCodeBench set it spends a median 17.7x fewer output tokens per problem than its own base at a cost of 4.6 points of pass@1 at matched greedy budget. This is a proof of concept: one LoRA run on a 10,205-example corpus. The costs are real and are reported next to the savings.
The reasoning channel is private by design. Users see normal code and normal tool calls; the compressed register never leaks into the output channel (0 of 46 completed problems in the measured run).
All numbers are paired measurements against the model's own re-measured base (Tesslate/OmniCoder-9B) on one pinned serving stack: vLLM 0.21.0, CUDA graphs, A100-80G, evalscope 1.9.0, temperature pinned per condition, 16,384-token generation budget, ceiling hits scored as failures.
Metric
Tessera
Base
Gap
LCB-1055 pass@1, greedy
34.9% [32.1, 37.8]
39.5% [36.6, 42.5]
−4.6 (95% CI [−7.5, −1.8])
LCB-1055 pass@1, temp 0.6
33.7%
45.9%
−12.2
IFEval-541 prompt-strict
40.9%
59.9%
−19.0 (95% CI [−24.5, −13.6])
Median output tokens per LCB problem
639
16,384 (at ceiling)
17.7x (median paired ratio)
Budget deaths at greedy (LCB)
21.6%
58.8%
The two models fail differently. The base almost never writes wrong code (95.9% of its completions pass) but thinks into the 16K ceiling on 58.8% of problems. Tessera completes 78% of problems at a median 639 tokens end to end and errs by writing wrong code. Forcing an empty think collapses accuracy from 66% to 4% on the archived 50-problem protocol: the compressed channel is load-bearing, not decoration.
Read the limitations before deploying. The instruction-following gap is large: the training corpus is 100% code-agentic with zero IFEval-style prompts, and the model falls into 16K reasoning loops on 46.8% of such prompts. Sampling does not help this model (greedy is the intended operating point). Whether the gaps are a data-coverage artifact or intrinsic to the compression is the successor's question; the paper argues coverage is the likely major cause and says what would prove it.
Usage
The repo ships vLLM-ready weights (LoRA merged, keys repacked, Qwen3.5-9B text config). Serve:
Pinned versions matter. All published numbers come from vLLM 0.21.0; vLLM 0.24.0 degraded this model in our validation and we do not recommend it. Greedy point estimates for this model family are stack-sensitive (the paper documents a 12-point spread across serving stacks), so treat scores measured on other stacks accordingly.
Two serving contracts to respect:
Pin temperature 0.0 for code. The model has no sampling headroom; the shipped generation_config.json is greedy by default.
Do not disable thinking.enable_thinking: false effectively disables the model (48 of 50 outputs empty in the ablation). There is no functional no-think mode.
The reasoning arrives on the standard reasoning channel (reasoning_content with the qwen3 parser) and the answer on content. You can discard the reasoning; it is not written for reading.
adapter/ contains the original LoRA (r=16 on Q/K/V/O) plus tokenizer and chat template, for anyone who wants to re-merge against the base or continue training. The base ships behind a vision-language wrapper, so a bare adapter merge needs a key repack; the merge and repack recipe is in the paper's Appendix B.
Training
LoRA SFT on 10,205 examples (7.62M target tokens, 2 epochs, one A100-80G, 6h58m). The corpus is 100% code-agentic: single-turn compressed-reasoning items, tool-call wrapped items, 1,500 execution-verified agentic trajectories compressed through tessera-compressor, and 400 multi-turn recall items. Training targets were rendered inference-faithfully (history turns carry empty thinks exactly as the serving template produces them). The run was gated by ten pre-registered behavioral probes frozen before any data existed.
License
Apache-2.0, same as the base model. Trained on permissively licensed data (per-record licenses listed in the paper's Appendix B).
