Instructions to use flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed")
model = AutoModelForMultimodalLM.from_pretrained("flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed

SGLang

How to use flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed with Docker Model Runner:
```
docker model run hf.co/flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

⚡ FlashModel-Qwen-3.6-35B-A3B-Maxed

A reasoning / coding / agentic / bilingual fine-tune of Qwen/Qwen3.6-35B-A3B.

Key capabilities:

Code: Rust, TypeScript, JavaScript, Python, QML, HTML, CSS, Bash
Controllable reasoning effort: put reasoning_effort=low|medium|high|xhigh|max in the system prompt to change reasoning depth — real control, not an artificial length cap
Agentic tool use: file read/write, shell, git operations
Bilingual EN/RU dialogue (≈22% RU in training)
MTP head preserved: mtp.* tensors kept from the original checkpoint so speculative decoding works

GGUF quants: flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed-GGUF

Training details

Method: LoRA (peft), merged into the base weights in this repo
LoRA: r=64, alpha=128, dropout=0.05
Targets (310 modules): attention q/k/v/o_proj, linear-attention in_proj_a/b/qkv/z + out_proj, shared-expert gate/up/down_proj (routed experts untouched — 85% of weights)
Data: 14.4k examples stratified:
- Reasoning: 6k (OpenR1-Math + simplescaling/s1K, 5 effort levels × ~1200 each)
- Code (local corpus): 4.9k (Rust 1101, QML 900, JS 900, Py 900, TS 711, CSS 250, HTML 66, Bash 100)
- Code (instruction): 800 (Magicoder-Evol-Instruct-110K, capped per language)
- Tool use: 1.5k (NousResearch Hermes + glaive)
- Bilingual chat: 1.2k (OpenAssistant oasst2, 21% RU)
Loss target: ~1.4 to avoid memorization; stopped at step 1500 of 2436 (train loss 0.43)
Context length in training: 16k tokens (longer examples were dropped, not truncated)
Effort data: reasoning traces regenerated at 5 budget levels so the marker maps to real behaviour
Hardware: single A100-80GB, 4-bit base + bf16 LoRA, ~4 h

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")

msgs = [
    {"role": "system", "content": "You are a coding and reasoning assistant.\nreasoning_effort=high\nAdjust the depth of your reasoning to the level given above."},
    {"role": "user", "content": "Напиши функцию бинарного поиска на Python."},
]
enc = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)
out = model.generate(enc["input_ids"].to(model.device), max_new_tokens=800)
print(tok.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=True))

Reasoning effort control

The effort level is read from the system prompt:

<system>
You are a coding and reasoning assistant.
reasoning_effort=medium
Adjust the depth of your reasoning to the level given above.
</system>

Value	Behaviour
`low`	1–2 sentences, minimal reasoning
`medium`	short paragraph
`high`	multi-paragraph reasoning
`xhigh`	long trace, explores alternatives
`max`	full chain-of-thought

This was trained with matched-pairs data (same question at 5 effort levels), so the mapping is behavioural rather than a truncation of the same text.

Known eval results (baseline → finetuned)

Metric	Baseline	Finetuned
Tool Use (8 scenarios)	6/8	7/8
Agentic (5 tasks)	5/5	5/5
Effort monotonic (low→max)	❌	✅
RU Perplexity	—	5.41

Full benchmark suite (HumanEval, MBPP, GSM8K, MMLU, SWE-bench) is planned.

Licence

Apache-2.0, inherited from Qwen/Qwen3.6-35B-A3B.

Downloads last month: 81

Safetensors

Model size

35B params

Tensor type

BF16

Model tree for flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed

Base model

Qwen/Qwen3.6-35B-A3B

Finetuned

(378)

this model

Quantizations

3 models

flashback2k
/

FlashModel-Qwen-3.6-35B-A3B-Maxed