Parameter	Value
direction_index	10.31
attn.o_proj.max_weight	1.15
attn.o_proj.max_weight_position	14.87
attn.o_proj.min_weight	0.72
attn.o_proj.min_weight_distance	12.89
mlp.down_proj.max_weight	1.27
mlp.down_proj.max_weight_position	14.41
mlp.down_proj.min_weight	0.10
mlp.down_proj.min_weight_distance	5.83

Performance

Metric	This model	Original model (empero-ai/Qwen3.8-2B-Distill)
KL divergence	0.0095	0 (by definition)
Refusals	4/100	83/100

Qwen3.8-2B

Developed by Empero

[!Note] This repository contains model weights and configuration files in the Hugging Face Transformers format.

These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, and other standard runtimes with Qwen3.5 architecture support.

Qwen3.8-2B is a full-parameter distillation of Qwen3.8 2.4T A95B into the Qwen3.5-2B architecture — the smallest member of the family, trained on the same curriculum as its larger siblings. The student saw ~30,000 curated teacher traces from our internal Qwen3.8 distillation datasets — dense chain-of-thought spanning mathematics, general reasoning, and instruction following, quality-filtered before training.

The objective: the same teacher, the same reasoning curriculum, in a model small enough for the edge. What changes across Qwen3.8-9B, Qwen3.8-4B and this model is the student's capacity — not the quality or the character of what it was taught.

Highlights

Distilled chain-of-thought — every answer opens with a <think> block learned directly from Qwen3.8 2.4T A95B traces rather than synthetic self-generated reasoning.
Same curriculum as the larger siblings — the same teacher and the same quality-filtered trace mix used for the 9B and 4B distills; only the student's capacity differs.
2B weight class — bf16 in ~4 GB; quantized builds run on phones, single-board computers, and CPU-only machines.
Native function calling per Qwen3.5's specification — no wrapper or tool-specific fine-tune required.
262,144-token native context, inherited from the Qwen3.5 base.
Full fine-tune — every parameter updated; not an adapter.

Model Overview

Type: Causal Language Model (text path of a vision-language base)
Base: Qwen/Qwen3.5-2B
Number of Parameters: 2B
Training: SFT (off-policy distillation) on ~30,000 teacher traces
Teacher: Qwen3.8 2.4T A95B (internal distillation datasets)
Context Length: 262,144 natively

Benchmark Results

Measured with lm-evaluation-harness, HF backend, identical settings for base and student. Both models are reasoning models and are evaluated with the CoT protocols (gsm8k_cot, mmlu_flan_cot_zeroshot); MMLU covers all 57 subjects (~1,700 questions). Flexible-extract is the primary metric; strict-match requires exact answer formatting.

Task	Metric	Qwen3.5-2B (base)	Qwen3.8-2B	Δ
gsm8k_cot	exact_match (flexible)	0.330	0.640	+0.310
gsm8k_cot	exact_match (strict)	0.545	0.640	+0.095
mmlu (CoT, 57 subjects)	acc (flexible-extract)	0.283	0.548	+0.265
mmlu (CoT, 57 subjects)	acc (strict-match)	0.004	0.225	+0.221

Sampling for generation: temperature=0.6, top_p=0.95, top_k=20 (Qwen3.5 recommended settings).

Quickstart

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "empero-ai/Qwen3.8-2B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "A snail is at the bottom of a 10-meter well. Each day it climbs 3 meters, each night it slips back 2. How many days until it escapes?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)

out = model.generate(inputs, max_new_tokens=16384,
                     temperature=0.6, top_p=0.95, top_k=20, do_sample=True)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

A recent transformers release with Qwen3.5 support is required, along with the Gated DeltaNet kernels (flash-linear-attention and a CUDA-matched causal_conv1d build) — without them the linear-attention layers fall back to slow, memory-hungry PyTorch ops.

Best Practices

Sampling: temperature=0.6, top_p=0.95, top_k=20. Greedy decoding on long generations is a known repetition-loop failure mode for reasoning models in this class.
Output length: allow generous max_new_tokens (16,384 recommended); every answer opens with a <think> block. Parse and strip the <think>...</think> span for end users.
Weight class, not curriculum: the reasoning form transfers from the same teacher traces the larger siblings learned from; what 2B parameters bound is capacity — factual recall and very hard multi-step problems. For harder workloads, step up to Qwen3.8-4B or Qwen3.8-9B.
Scope: the trace mix emphasizes mathematics, reasoning, and instruction following; the 9B additionally trains on code, so use Qwen3.8-9B for code-heavy workloads. The fine-tune is text-only; vision behavior is inherited from the base and was not evaluated here.

Stay in the loop

Support / Donate

If this model helped you, consider supporting the project:

BTC: bc1qx6zepu6sfkvshgdmc4ewu6pk6rpadvpgffpp7v
LTC: ltc1qv2mefzps2vtjcpwfx8xxdrpplrcvltswm68r7x

Provenance & licensing

Weights are released under Apache-2.0, inherited from the Qwen3.5-2B base. Shared for research and experimentation, as-is.

Acknowledgements

Developed and released by Empero
Base model: Qwen3.5-2B (Alibaba Qwen team)
Training: TRL + Transformers
Linear-attention kernels: flash-linear-attention, causal_conv1d
Evaluation: lm-evaluation-harness (EleutherAI)

MihaiPopa-1/Qwen3.8-2B-Heretic-Balanced-LoRA

This is a decensored version of empero-ai/Qwen3.8-2B-Distill, made using Heretic v1.4.0

Abliteration parameters