mconcat/Qwopus3.6-27B-v2-AWQ-4bit - Fast, Reliable, and Scalable Inference on FriendliAI

Quick start

Requires vLLM ≥ 0.21.0:

bash
vllm serve mconcat/Qwopus3.6-27B-v2-AWQ-4bit \
  --tensor-parallel-size 1 \
  --max-model-len 16384 \
  --speculative-config '{"method": "mtp", "num_speculative_tokens": 3}' \
  --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --trust-remote-code

Benchmarks

Evaluated with lm-evaluation-harness on a single NVIDIA B300 SXM6, 100 samples per task, 0-shot CoT, max_gen_toks=4096:

Task	Qwen 3.6 27B (base)	Qwopus 3.6 v2 (source BF16)	This (AWQ-4bit)
GSM8K (flexible-extract)	65.0%	87.0%	85.0%
ARC-Challenge (acc_norm)	46.0%	45.0%	47.0%
TruthfulQA-MC2	55.1%	59.3%	59.3%
IFEval (inst_level_strict)	40.5%	42.3%	42.9%

Quantization preserves accuracy within standard error of the BF16 source on every task, and matches the source on TruthfulQA. The Claude Opus reasoning gain over the Qwen 3.6 base (+20 pp on GSM8K) is retained.

Throughput

Measured on a single NVIDIA B300 SXM6 with vLLM 0.21.0 and torch.compile enabled:

Setup	Throughput	Speedup
Batch = 1, no MTP	115 tok/s	1.00×
Batch = 1, MTP `num_speculative_tokens = 3`	251 tok/s	2.19×
Batch = 8 continuous batching, no MTP	880 tok/s	—

MTP speculative decoding hits an Avg Draft acceptance rate of ~77 % (per-position: 0.92 / 0.79 / 0.65) with a mean acceptance length of ~3.3, measured on a mixed reasoning + code prompt set at greedy decoding.

Self-test of tool calling with --tool-call-parser qwen3_coder: passes (model emits well-formed <tool_call>...</tool_call> syntax).

Quantization

Precision	Modules
INT4 asymmetric, group_size = 128	`q_proj`, `k_proj`, `v_proj`, MLP `gate_proj`, MLP `up_proj`, DeltaNet `in_proj_qkv`, `in_proj_z`
BF16	`o_proj`, MLP `down_proj`, `lm_head`, `embed_tokens`, norms, DeltaNet small projections (`in_proj_a`, `in_proj_b`), DeltaNet `out_proj`, vision tower, multimodal projector, 1-layer MTP head, first 5 and last 5 transformer layers (APEX edge protection)

The AWQ skip list also names every mtp.* linear module explicitly so the MTP draft head stays unquantized — previous revisions of this checkpoint omitted those entries, which caused vLLM to build the MTP head with AWQ-packed parameters and produced 0 % draft acceptance.

Tuned with AutoRound on 1024 self-generated reasoning traces (200 iterations per block, batch_size=1).

Calibration data: 1024 self-generated traces from the BF16 source model (256 prompts × 4 generations) covering math, code, logic, analysis, creative writing, general knowledge, tool calling, and Korean.

Files

File	Size	Purpose
`model-*.safetensors` (13 shards)	~25 GB	Main quantized weights
`model_extra_tensors.safetensors`	~1 GB	MTP head + edge-protected layers (BF16)
`quantization_config.json`	<1 KB	AWQ config (`quant_method=awq`, `bits=4`, `group_size=128`, `zero_point=true`) with BF16 MTP skip entries
`config.json` + tokenizer + processor configs	<100 MB	Standard metadata

Total checkpoint size: ~26 GB (down from ~54 GB BF16 source).

License

Apache 2.0 (inherited from the base model).

Qwopus3.6-27B-v2-AWQ-4bit

Get help setting up a custom Dedicated Endpoints.

README