Blox AI Qwen 3 1.7B Instruct RKLLM W8A8 (v1)

Qwen 3 1.7B with thinking mode, RKLLM-quantized W8A8 for RK3588 NPU. First Qwen 3 fine-tune of the Blox AI assistant.

Training: 2 epochs LoRA (rank 16, alpha 32) on 174 synthetic examples, effective batch 4, final loss 0.39. Smoke-tested locally for Kubernetes hallucination correction (rule 11), false-positive restart resistance (rule 8), null-action verdicts (rule: don't invent fixes).

Requires librkllmrt.so v1.2.3 on-device (the prior 1.1.4 runtime did not support Qwen3ForCausalLM). functionland/blox-ai:release image rebuilt 2026-05-27 with v1.2.3.

File size: 2,375,021,452 bytes (2.21 GB). Split into 2 chunks because GitHub release assets cap at 2 GiB per file. download_model.sh reassembles + SHA-verifies.

SHA-256 of assembled file:
8843d4612d42a71605c8d0b38cf9a758a5a53f65a8dde3f17f9d4549b9794e87

Blox AI smaller model — Qwen 2.5 1.5B Instruct RKLLM W8A8 (v1)

Smaller model variant for resource-constrained Blox AI deployments.

Qwen 2.5 1.5B Instruct, RKLLM-quantized W8A8 for RK3588 (sourced from
c01zaut/Qwen2.5-1.5B-Instruct-RK3588-1.1.4 on HuggingFace, variant
opt-1-hybrid-ratio-0.5 for the lowest perplexity at this size).

~1.89 GB → loads into ~1.9 GB shared memory at runtime. Half the
RAM footprint of the 3B W8A8 model (~3 GB shmem) so it fits in the
~3 GB RAM headroom typically available on a 7.7 GB RK3588 board
that's also running the full Fula stack (kubo + cluster + go-fula +
pinning + gateway + fxsupport). The 3B model OOM-killed during cold
start on tight-RAM scenarios; this 1.5B variant avoids that.

Tool-calling supported. For the Blox AI runbook-guided diagnostics
use case the smaller model is enough; runbook gives explicit
recipes so the model is mostly classifying user symptoms and
calling tools, not doing open-ended reasoning.

Integrity

File	Size (bytes)	SHA-256
qwen2.5-1.5b-instruct-rk3588-w8a8.rkllm	2,035,400,284	b09198d0b389615edfea0def0032722fc853e1d90ccc47ab6c545f8568af8a13

Download

Single file, no chunking (under the 2 GiB GitHub Release asset cap):

wget https://github.com/functionland/blox-ai/releases/download/model-qwen-2.5-1.5b-w8a8-v1/qwen2.5-1.5b-instruct-rk3588-w8a8.rkllm
sha256sum qwen2.5-1.5b-instruct-rk3588-w8a8.rkllm
# Expected: b09198d0b389615edfea0def0032722fc853e1d90ccc47ab6c545f8568af8a13

Blox AI base model — Qwen 2.5 3B Instruct RKLLM W8A8 (v1)

Base model for Blox AI v1 — Qwen 2.5 3B Instruct, RKLLM-quantized W8A8 for RK3588.

Source: c01zaut/Qwen2.5-3B-Instruct-rk3588-1.1.1 on HuggingFace, RKLLM toolkit 1.1.1 W8A8.

The 3.74 GB model is split into two chunks because GitHub Release assets are capped at 2 GiB per file. download_model.sh fetches both, concatenates them in order (chunk-aa then chunk-ab), and verifies the assembled SHA against the value below.

Integrity

File	Size (bytes)	SHA-256
chunk-aa	1,992,294,400	2b14f9f9534bc5b755901c60c329beb5ac4e3dc0ff8e1ca29072a1cc313d4f7e
chunk-ab	1,746,052,348	cbb0f6d43af94d9a2cde0c164810a9b91c6d2c03b38637119c63409b96d2f046
assembled	3,738,346,748	b7cf8b1c10140ac380535a52602d2ecc862aa96a84e3cf5d8267b6e54cca2607

Manual reassembly

wget https://github.com/functionland/blox-ai/releases/download/model-qwen-2.5-3b-w8a8-v1/chunk-aa
wget https://github.com/functionland/blox-ai/releases/download/model-qwen-2.5-3b-w8a8-v1/chunk-ab
cat chunk-aa chunk-ab > qwen2.5-3b-instruct-rk3588-w8a8.rkllm
sha256sum qwen2.5-3b-instruct-rk3588-w8a8.rkllm
# Expected: b7cf8b1c10140ac380535a52602d2ecc862aa96a84e3cf5d8267b6e54cca2607