Chimera-1

Model details

Chimera-1 asks one question: can two specialists be combined into a single model without training at the merge step, and without losing general ability? We made two light specialists from Qwen2.5-3B-Instruct, one for Roman-numeral conversion and one for character-level word operations. Then we merged them with TIES, a published merging method. The merge runs on a CPU in minutes, and the result is one 3B model with the same inference cost as its base.

Status of claims. The merge results on this page were measured by Tetracta, separately from Model X-Ray. They have not been reviewed externally. Any earlier claim that Model X-Ray guided, predicted or explained this merge is withdrawn. See the correction record.

Intended use

Primary uses

• Research on model merging: what a training-free merge can and cannot carry into one model.
• Checking or extending our measurements.
• Trying the two implanted skills side by side with the base model.

Out of scope

• Commercial use of any kind. The licence does not allow it.
• Use as a production assistant. This is a method demo.
• Relying on its Roman-numeral or letter-operation answers without checking them.
• Merging with models that do not share its base. Weight merging needs a shared base.

Training data

• Base model: Qwen2.5-3B-Instruct, trained by its own developers on data they document. We did not change its pre-training.
• Specialists: light fine-tunes on synthetic, verifiable task data that can be generated without limit. One task is Roman-numeral conversion. The other is character-level word operations, such as spelling a word backwards. We checked each specialist on its task before merging. We do not publish the fine-tuning details.
• Merge step: uses no data.
• Evaluation sets: held-out skill items, MMLU, ARC-Challenge, HellaSwag, and Belebele in five languages.

Training procedure

Evaluation

Pilot: 80 held-out items per skill

Head-to-head against the base: 1,210 questions

Both models ran on the same log-likelihood harness. As a sanity check, the base scores 67.0% on MMLU, consistent with its publicly reported value.

The two target skills rose 8.8 and 25.0 points. All eight standard benchmarks stayed inside the ±2-point band, with a largest change of 1.3 points. This supports one bounded statement: on these 1,050 general questions we detected no regression larger than run-to-run noise. It does not show that nothing else changed.

The pilot and the head-to-head use different item sets and scoring, so their skill numbers differ, for example 89% and 91.2% on Roman numerals.

Controls and failures

The same campaign ran about 10 merge experiments at 0.5B and 3B, within one model family and across families. Some failed, and the failures are part of the result.

• No headroom: our first 3B attempt used tasks the base already solved (100% and 83%). The merge had nothing to add.
• No complementarity: merging a generalist with a coder that was no better at code diluted both, by 8 points on code and 11 on multilingual tasks.
• Too far from the base: aggressively merging a heavily continued-pretrained coder gave 0% on code and chance on everything else. This is our cleanest negative control.
• Three exploratory merge variants did not beat plain TIES on average.

Across these experiments, outcomes were associated with three conditions: headroom on the task, complementary parents, and parents still close to the shared base. We treat that as a working hypothesis for this campaign, not a validated rule.

Limitations and what is not done yet

• It is a method demo. The two skills are deliberately simple and verifiable, chosen because the base was weak at them.
• Tetracta evaluated it. No external review or reproduction has been done.
• Item counts are small: 75 to 200 per benchmark and 80 per skill. There are no confidence intervals, only the ±2-point noise band.
• One base model, one merge method and one pair of skills. The result may not hold for other bases or harder skills.
• Both skills still make errors after the merge: 91.2% and 62.5% in the head-to-head. Letter operations stayed below specialist B's 75% in the pilot.
• The three-condition pattern above comes from about 10 experiments. It is not a law.

Not done yet: an external review, confidence intervals, a safety evaluation of the merged model, and tests on harder or real-world skills.

Risks and safety

Chimera-1 inherits the behaviour, biases and safety tuning of Qwen2.5-3B-Instruct. We have not measured whether the merge changed its safety behaviour, and no safety evaluation of Chimera-1 is on record.

The licence allows non-commercial research use only. Do not use the weights in a product, a commercial demo or paid work.

The "no detected regression" statement covers the listed checks only. Do not read it as a guarantee about other tasks.

Environmental impact

All of the roughly 10 experiments in the campaign ran on two local consumer GPUs, with no cloud compute. The merge itself runs on a CPU in minutes. We did not record GPU-hours, and we have not measured energy use or emissions.

How to access

Weights and card are on Hugging Face: huggingface.co/tetracta/Chimera-1-Qwen2.5-3B. The model loads like its base. This snippet is from the Hugging Face card:

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("tetracta/Chimera-1-Qwen2.5-3B", dtype="bfloat16")
t = AutoTokenizer.from_pretrained("tetracta/Chimera-1-Qwen2.5-3B")
# try: "Convert the Roman numeral CDXLIV to a regular number." / "Spell the word 'measure' backwards."

The full study, including the failed variants, is in our note Chimera-1: two skills, one model, training-free.

Citation

@misc{tetracta_chimera1_2026,
  title  = {Chimera-1: two skills, one model, training-free merge (method demo)},
  author = {Tetracta},
  year   = {2026},
  url    = {https://huggingface.co/tetracta/Chimera-1-Qwen2.5-3B}
}

The Hugging Face card has no BibTeX entry. This is the one we suggest.

Changelog