What is actually new in Dhenu 2

The Dhenu2 family shipped last month: 1B, 3B, and 8B variants built on Llama 3.1 and 3.2, trained on 1.5 million instructions across more than 4,000 agricultural categories. There is a developer platform alongside them so teams can build on the models. The cards are on Hugging Face, and deployment to KissanAI partners is rolling out over the next few weeks.

The release notes cover all of that. I want to stay with the climate work because it is the part I care about most, and because a press note flattens it into a feature bullet.

Here is what changed under the hood.

Climate-state context is a first-class input

A year ago Dhenu 1.0 answered questions about agriculture. It did not know what week it was, what the monsoon had done, or where the farmer was standing. The advisory was generic enough that a farmer in Marathwada and a farmer in north Bihar would get similar answers to similar questions.

Dhenu 2 takes a climate-state vector as part of every inference. That vector includes the current week of the agricultural calendar, the rolling 30-day rainfall departure from normal for the district, the short-range IMD forecast for the next 7 days, the current soil-moisture state at three depths where the sensor coverage exists, and a coarse drought-stress index. The model was continued-pre-trained on a corpus of climate-conditioned advisories and instruction-tuned on 92,000 pairs where the same agricultural question gets a different correct answer based on the climate state at the time of asking.

When a farmer asks “should I sow now,” the answer in a sub-normal-rainfall year should differ from the answer in a normal year. Dhenu 2 now receives the context required to make that distinction.

Climate state is also the dimension where generic models will be slowest to catch up, because it is not sitting in a scraped corpus waiting to be trained on. It is a live, perishable, district-by-district feed that has to be plumbed into inference week after week, and the durable advantages in applied AI are increasingly built out of exactly this kind of unglamorous data engineering. Medicine figured this out long ago: a diagnosis without the patient’s current vitals is a textbook entry, not a consultation. Agriculture is only now getting its vitals wired in.

The long tail of crops actually shows up

Dhenu 1.0 was strong on the major commodity crops and weaker on horticulture and the regional minor cereals. We measured 4.31 on a held-out eval that was 64% weighted to rice, wheat, cotton, sugarcane, soybean, and maize.

For Dhenu 2 we rebalanced. The new training corpus has 38% of its instruction pairs on horticulture (mango, pomegranate, banana, vegetables) and on the minor millets that are coming back as climate-resilient alternates (ragi, jowar, bajra, kodo). We hand-built a 14,000-pair set on pomegranate alone, working with horticulture extension publications from Maharashtra and Karnataka, because we were getting too many wrong answers about a crop that millions of small growers depend on.

The eval result was a mean score of 4.47 on the same rubric. The variance across crops also dropped. Dhenu 1.0 had a 0.61-point gap between its best-served crop (rice) and its worst-served (pomegranate). Dhenu 2 has a 0.18-point gap.

Image inputs for crop protection

This is the big architecture change. Dhenu 2 has a vision encoder. Farmers can send a photograph of a leaf, a pest, a stand of plants, and the model can interpret what it is looking at and return a contextualised diagnostic.

The vision stack is a CLIP-style encoder trained on the image corpus we have been building through the year: pest specimens labelled at species level, disease symptoms across stages of progression, soil-deficiency leaf colouration, weed identification across regional varieties. The encoder feeds into the same Dhenu language backbone via a learned projection; the model card carries the corpus details.

What this gets us, practically, is that a farmer in Pune who sends a photograph of a pomegranate leaf with early-stage bacterial blight does not need to know the word “blight.” She sends the photo, the model identifies the symptom, asks two or three clarifying questions in her language, and returns an action plan with timing and severity-conditioned recommendations.

The vision model is the right tool for crop protection in a country where farmers do not know the disease names in the language the literature is written in, and where extension officers are too few to visit every field that is asking for help.

What is still hard

Long multi-turn dialogue where the farmer reveals context over four or five turns is improved but not solved. The model occasionally still re-asks for information the farmer provided two turns earlier. This is on us to fix, partly through better dialogue-state tracking, partly through preference data we will collect in the next quarter.

Region-specific seed varieties and brand-level pesticide recommendations remain conservative. We will not have the model name a brand it has not been trained to be confident about, because the cost of a wrong brand recommendation in a real spray is a season. The right product discovery layer for branded inputs is on the platform side, not in the foundation model. We will write about that separately.

Where it is

Dhenu 2 weights, tokenizer, eval set, and the vision encoder are at huggingface.co/KissanAI under the Dhenu2 family. The model card has the training config and the corpus statistics. The eval prompts and methodology are open. Beat us on them.