empero-ai

Qwen3.8-35B-A3B-Distill

Deploy Dedicated

README

License: apache-2.0

Highlights

Distilled chain-of-thought — every answer opens with a <think> block learned directly from Qwen3.8 teacher traces rather than synthetic self-generated reasoning.
Mathematics and code emphasis — the trace mix is deliberately weighted toward hard math and competitive programming, the domains where distillation moves the needle most at this scale.
Sparse MoE efficiency — 35B total parameters, ~3B active per token; 256 experts with 8 routed per token.
Attention and experts both adapted — our internal MoE training pipeline updates the attention path and the routed and shared expert stacks, not just attention.
Native function calling per Qwen3.6's specification — no wrapper or tool-specific fine-tune required.
262,144-token native context, inherited from the Qwen3.6 base.

Model Overview

Type: Causal Language Model (text path of a vision-language base)
Base: Qwen/Qwen3.6-35B-A3B
Number of Parameters: 35B total / ~3B active per token
Architecture: 40 layers, 256 experts, 8 experts per token, hybrid linear + full attention
Training: SFT (off-policy distillation) on curated teacher traces via our internal MoE training pipeline
Teachers: Qwen3.8 2.4T A95B and Qwen3.8 Flash Next (internal distillation datasets)
Context Length: 262,144 natively

Benchmark Results

Measured with lm-evaluation-harness, HF backend, bfloat16, identical settings and seed for base and student. Zero-shot, loglikelihood scoring.

Table with columns: Task, Metric, Qwen3.6-35B-A3B (base), Qwen3.8-35B-A3B, Δ
Task	Metric	Qwen3.6-35B-A3B (base)	Qwen3.8-35B-A3B	Δ
MMLU (57 subjects)	acc	0.838	0.834	−0.004
ARC-Challenge	acc	0.548	0.582	+0.034
ARC-Challenge	acc_norm	0.548	0.591

The MMLU difference is within noise (standard error 0.003 on each measurement). The ARC gains are outside it.

Quickstart

python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "empero-ai/Qwen3.8-35B-A3B-Distilled"
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", return_dict=True).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["input_ids"].shape[1]:], skip_special_tokens=True))

A recent transformers release with Qwen3.6 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.

AutoModelForCausalLM loads the text path (34.7B parameters). The vision tower is retained in the checkpoint and is reachable via AutoModelForImageTextToText.

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.
Scope: the model learned from teacher traces, not from its own rollouts — it inherits the teachers' reasoning style, including occasional over-long deliberation on easy questions.

Limitations

Shorter responses. The student was trained on 8,192-token examples and produces noticeably shorter outputs than the base. Long chains of thought are more likely to be cut short, so behaviour on long-form generation and long-context workloads may be degraded relative to the base.
A v2 is in training with longer-context support, aimed squarely at the point above.
Vision is untouched. The fine-tune is text-only; vision behaviour is inherited from the base and was not evaluated.

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 Qwen3.6-35B-A3B base. Shared for research and experimentation, as-is.

Acknowledgements

Developed and released by Empero
Base model: Qwen3.6-35B-A3B (Alibaba Qwen team)
Teachers: Qwen3.8 2.4T A95B and Qwen3.8 Flash Next
Training: TRL + Transformers
Linear-attention kernels: flash-linear-attention, causal_conv1d
Evaluation: lm-evaluation-harness (EleutherAI)
— a portion of the prompts used to generate this model's synthetic training data came from those interactions.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

empero-ai

Model Tree

Base

Qwen/Qwen3.6-35B-A3B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality