---
license: agpl-3.0
base_model: lordx64/Qwable-v1
base_model_relation: quantized
pipeline_tag: text-generation
library_name: transformers
tags:
  - awq
  - int4
  - w4a16
  - compressed-tensors
  - llm-compressor
  - quantized
  - moe
  - qwen3.6
  - chain-of-thought
  - agentic
  - tool-use
  - vllm
language:
  - en
---

# Qwable-v1-AWQ

AWQ 4-bit (W4A16) quantization of [lordx64/Qwable-v1](https://huggingface.co/lordx64/Qwable-v1) — a **35B-total /
3B-active text generation** Mixture-of-Experts model (`Qwen3_5MoeForConditionalGeneration`, Qwen3.6
family, with hybrid linear / full attention). Per the base model card it is **text-only** and aimed at
reasoning, agentic tool-use, and coding (see [Capabilities](#capabilities)).

**Variant**: AWQ weight-only (W4A16) — int4 symmetric weights, group size 128, activation-aware scaling; activations stay BF16
**Disk size**: ~22 GB (vs ~72 GB BF16, ~3.3×)
**Quantized by**: [sahilchachra](https://huggingface.co/sahilchachra)
**Tooling**: `llm-compressor` AWQ (`oneshot`) — activation-aware, calibrated on general instruct chat (UltraChat-200k)

> **Note on what is quantized**: only the linear weights that hold the bulk of the parameters are
> taken to int4 — the 256-way routed experts, the shared experts, and the full-attention
> projections. The linear/Gated-Delta-Net (mamba-style) layers, the MoE routers, embeddings,
> `lm_head`, the MTP head and all norms are kept in BF16 for stability. The architecture also carries
> a vision tower (`Qwen3_5MoeForConditionalGeneration`), which is likewise kept in BF16 — but the base
> model is documented as text-only, so this quantization neither adds nor validates any image
> capability. The headline variant name reflects the dominant (expert/attention) quantization; the
> on-disk size averages the int4 and BF16 halves of the model.

## Capabilities

Unchanged from the base model — quantization only changes weight precision, not behavior. Per the
[base model card](https://huggingface.co/lordx64/Qwable-v1):

- **Reasoning** — thinks in explicit `<think>…</think>` chains-of-thought.
- **Agentic tool-use** — emits `<tool_use>` XML blocks for file/shell operations (activates with
  agent-style system prompts or prior `<tool_result>` turns).
- **Coding** — designed for agentic coding tasks with multi-turn agent interactions.
- **Context length**: 4096 tokens (training) / 16384 tokens (serving).

See the base card for limitations (narrow training distribution, tool-name differences, reasoning
inherited from the Opus-4.7 distill).

## Smoke test

Loaded and run with **transformers** on an NVIDIA Thor (Blackwell) device. The model loads, runs the
hybrid linear-attention + int4 MoE path, and produces coherent text from a chat-templated prompt. A
structure census confirms only the intended decoder Linears are int4 (routed experts, shared expert,
full-attention `q/k/v/o`) with the routers, linear-attention, vision, MTP and norms left in BF16. This
is a functional smoke test only — it is **not** a quality benchmark.

### Test device

- **GPU**: NVIDIA Thor (Blackwell)
- **CPU / memory**: 14-core ARM (aarch64), 122 GB unified memory
- **Software**: JetPack / L4T R38.4 (Ubuntu 24.04), CUDA 13.0, driver 580, kernel 6.8.12-tegra

## What's quantized

| Quantized → int4 (AWQ W4A16) | Kept in BF16 |
|---|---|
| Routed experts (`mlp.experts.*.{gate,up,down}_proj`, 40 layers × 256 experts) | Linear / Gated-Delta-Net layers (`*.linear_attn.*`) |
| Shared experts (`mlp.shared_expert.{gate,up,down}_proj`) | MoE routers (`mlp.gate`), shared-expert gates |
| Full-attention projections (`self_attn.{q,k,v,o}_proj`) | Embeddings, `lm_head`, MTP head, all norms |
| | Vision tower (`model.visual.*`) — present in the arch, unused for text |

## Usage (vLLM)

```python
from vllm import LLM, SamplingParams

llm = LLM(model="sahilchachra/Qwable-v1-AWQ", dtype="bfloat16", max_model_len=16384, trust_remote_code=True)
out = llm.generate(["Hello!"], SamplingParams(temperature=0.7, top_p=0.9, max_tokens=128))
print(out[0].outputs[0].text)
```

Runs on GPUs with `compressed-tensors` W4A16 support (vLLM unpacks the int4 weights for you).

## Notes

- Weight-only AWQ (W4A16): weights are int4 (group size 128, symmetric, activation-aware scales),
  activations remain BF16.
- Format: `pack-quantized` (compressed-tensors), per-expert layout — the standard layout vLLM consumes
  for quantized MoE.
- Loading requires `compressed-tensors` and a recent `transformers` (the `qwen3_5_moe` architecture).
- Smoke-tested only; not formally benchmarked for quality.
- Sibling quantization: [sahilchachra/Qwable-v1-NVFP4A16](https://huggingface.co/sahilchachra/Qwable-v1-NVFP4A16) (NVFP4 for Blackwell GPUs).

## Original model

See [lordx64/Qwable-v1](https://huggingface.co/lordx64/Qwable-v1) for full lineage, intended use, and
limitations. License (AGPL-3.0) is inherited from the base model.