Kimi K2.6 Bootstrap

moonshotai/Kimi-K2.6 has passed bootstrap and is active in Proof of Compute on Gonka mainnet. The timeline and transaction examples below remain useful for understanding how activation worked and for operations such as delegation; for current deployment defaults (including node-config.json), see the Host Quickstart.

Epoch 251 was when the model group moonshotai/Kimi-K2.6 became eligible.

This document explains how to minimize the chance of weight reductions, whether or not the model gets enough participants.

Note

The bootstrap can take multiple epochs, depending on how many participants are ready. Before activation, no weight reduction happens if participants submit their choice explicitly and hosts who are going to deploy submit PoCIntent.

Timeline

1. Before block `3873996`, all participants must submit:

- `PoCIntent` - if they are going to deploy `Kimi-K2.6`. Hosts should keep nodes deployed with `Qwen235B` and switch only after evaluation at block `3873996`
- `PoCDelegation` / `PoCRefusal` - if they are NOT going to deploy `Kimi-K2.6`

2. At block `3873996`, the chain runs pre-evaluation to check whether it should try to activate the model based on `PoCIntent` / `PoCDelegation`

- If the model becomes pre-eligible => hosts who submitted `PoCIntent` should switch their model nodes to `Kimi-K2.6` (there are no CPoC in this 500-block window)
- If the model does not become pre-eligible => hosts who submitted `PoCIntent` should keep their nodes on `Qwen235B`

3. At block `3874496`, PoC starts

Possible Scenarios

The bootstrap of a new model can follow these main scenarios:

The model does not pass the pre-evaluation at block 3873996 and does not become eligible
Everyone who submitted PoCIntent keeps their full weight (no punishment)
Everyone who submitted PoCDelegation / PoCRefusal keeps their full weight (no punishment)
Everyone who submitted nothing loses 15% of their weight

=> it is important to explicitly send a transaction with your intended behavior

The model passes the pre-evaluation at block 3873996 but does not become eligible at PoC
Everyone who participated in PoC keeps their full weight from Qwen235 (no punishment)
Everyone who submitted PoCDelegation / PoCRefusal keeps their full weight (no punishment)
Everyone who submitted nothing loses 15% of their weight
Everyone who submitted PoCIntent and does not participate loses 15% of their weight

If the model passes both checks, punishment follows the usual scenarios described in the documentation.

Instructions for hosts who are going to deploy Kimi-K2.6

1. Send `PoCIntent` to the chain:

export NODE=https://node3.gonka.ai/
./inferenced tx inference declare-poc-intent moonshotai/Kimi-K2.6 \
  --from node-2 \
  --node "$NODE" \
  --chain-id gonka-mainnet \
  --keyring-backend file \
  --gas auto \
  --gas-adjustment 1.3 \
  -y \

2. Check your setup and make sure the `Kimi-K2.6` weights are downloaded and you can deploy the model successfully

Wait for block 3873996+ and check whether the model becomes pre-eligible:

NODE=https://node3.gonka.ai
MODEL='moonshotai/Kimi-K2.6'

HEIGHT=$(curl -sG "$NODE/chain-rpc/block_search" \
  --data-urlencode "query=\"bootstrap_model_preeligibility.model_id='$MODEL'\"" \
  | jq -r '[.result.blocks[].block.header.height|tonumber]|max')

echo "Latest snapshot at height $HEIGHT"

curl -s "$NODE/chain-rpc/block_results?height=$HEIGHT" \
  | jq --arg m "$MODEL" '
      .result.finalize_block_events[]
      | select(.type=="bootstrap_model_preeligibility")
      | (.attributes | from_entries) as $a
      | select($a.model_id==$m)
      | $a'

The results will be sent in all channels.

4. Switch the model to Kimi-K2.6 if needed

Example command to deploy Kimi-K2.6 on 4xB200 / 8xB200:

curl -X POST http://localhost:9200/admin/v1/nodes \
     -H "Content-Type: application/json" \
     -d '{
       "id": "<NODE_ID>",
       "host": "<NODE_IP>",
       "inference_port": 5050,
       "poc_port": 8080,
       "max_concurrent": 500,
       "models": {
         "moonshotai/Kimi-K2.6": {
           "args": [
             "--tensor-parallel-size", "4",
             "--enable-expert-parallel",
             "--trust-remote-code",
             "--mm-encoder-tp-mode", "data",
             "--tool-call-parser", "kimi_k2",
             "--reasoning-parser", "kimi_k2",
             "--attention-backend", "FLASHINFER_MLA",
             "--disable-custom-all-reduce",
             "--gpu-memory-utilization", "0.95",
             "--max-num-seqs", "128",
             "--max-model-len", "240000"
           ]
         }
       }
     }'

Instructions for hosts who are NOT going to deploy Kimi-K2.6

1. Check if you trust any host who is going to deploy Kimi K2.6 / send `PoCIntent`

Current intents:

import requests

NODE = "https://node3.gonka.ai"
MODEL = "moonshotai/Kimi-K2.6"

participants = requests.get(f"{NODE}/v1/epochs/current/participants").json()["active_participants"]["participants"]

intents = []
for p in participants:
    addr, weight = p["index"], int(p["weight"])
    resp = requests.get(f"{NODE}/chain-api/productscience/inference/inference/poc_delegation/{addr}").json()
    for i in resp.get("intents") or []:
        if i["model_id"] == MODEL:
            intents.append((addr, weight))

total = sum(int(p["weight"]) for p in participants)
intent_weight = sum(w for _, w in intents)

print("Intent from:")
for addr, weight in intents:
    print(f"{addr} : {weight}")
print()
print(f"Intent weight: {intent_weight} / {total}")

2. Send delegation or refusal

Delegation:

export NODE=https://node3.gonka.ai/chain-rpc/
./inferenced tx inference set-poc-delegation moonshotai/Kimi-K2.6 <DELEGATEE> \
  --from gonka-account-key \
  --node "$NODE" \
  --chain-id gonka-mainnet \
  --keyring-backend file \
  --gas auto \
  --gas-adjustment 1.3 \
  -y

Refusal:

export NODE=https://node3.gonka.ai/chain-rpc/
./inferenced tx inference refuse-poc-delegation moonshotai/Kimi-K2.6 \
  --from gonka-account-key \
  --node "$NODE" \
  --chain-id gonka-mainnet \
  --keyring-backend file \
  --gas auto \
  --gas-adjustment 1.3 \
  -y