14.8B (dense; 14,822,046,720 parameters)
40,960 tokens
Apache-2.0
Open-weight biological language model derived from Qwen3-14B and specialized for joint reasoning over natural language and raw DNA. Carbon-X retains Qwen3's dense 40-layer transformer backbone while introducing a HybridDNATokenizer that combines the original Qwen vocabulary with 5,515 DNA-specific tokens, including all 4,096 six-base combinations, shorter tail tokens, DNA boundary markers, and an out-of-vocabulary DNA token; one DNA token can therefore represent six nucleotide bases. Training began with approximately 10B embedding-only tokens while the transformer was frozen, followed by 302B mid-training tokens using a mixture of 55% raw DNA, 20% English, 15% biological text, and 10% instruction data, then three supervised-finetuning epochs and two rounds of asynchronous GRPO with verifiable rewards. The publisher reports a 66.7 average across five DNA benchmarks without reasoning, including 74.9 on BioReason Coding, 85.1 on BioReason Non-SNV, 63.6 on GenomeQA, 39.4 on NextBase, and 70.4 on FrontierVEP, alongside 79.3 prompt-level strict accuracy on IFEval. The released BF16 checkpoint contains approximately 29.64 GB of weights. Official GGUF variants include BF16, Q8_0, Q6_K, Q5_K_M, and Q4_K_M files from approximately 29.65 GB down to 9.64 GB; the publisher recommends Q8_0 and reports near-BF16 biological benchmark performance. The custom tokenizer requires llama.cpp b11742 or newer because older versions can load the model while silently tokenizing DNA incorrectly. Thinking mode is supported but was disabled during post-training and reported evaluations. Carbon-X is a research model rather than a clinically validated medical system and may hallucinate biological functions or make sequence-counting and editing errors.