Instructions to use AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus with Transformers:

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

pipe = pipeline("text-generation", model="AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus

SGLang

How to use AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus 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 "AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus with Docker Model Runner:
```
docker model run hf.co/AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus
```

DeepSeek-V4.1-Flash-AZ-Focus

DeepSeek-V4.1-Flash with the second-guessing taken out: about half the thinking, and agentic coding at parity.

DeepSeek-V4.1-Flash is a strong reasoner, but much of its thinking is hesitation: "Wait…", "Hmm…", "Let me double-check…", re-deriving answers it already had. We optimised the model's weights against that habit, while leaving alone the checking it does when it works as an agent (running tests, inspecting output before it submits). AZ-Focus still reasons step by step. It just doesn't loop.

About half the thinking tokens at every reasoning-effort setting, with general accuracy close to the original.
Agentic coding at parity: SWE-bench Verified 98.5 vs 99 (first 100 tasks, two runs each).
Beyond the dial. At the lowest effort setting AZ-Focus needs a third to a half less thinking than stock DeepSeek-V4.1-Flash's own minimum, and it combines with the dial.
Drop-in replacement. Same architecture, tokenizer, chat template, file layout, size and MXFP4 / FP8 formats as the original. Serve it with the same engine and settings; no adapter, patch or extra code.

Made by AxonZeta, an independent AI lab in Manchester, UK.

Agentic coding

Benchmark	DeepSeek-V4.1-Flash	AZ-Focus
SWE-bench Verified, first 100 tasks (mini-SWE-agent, two runs)	99 / 99	99 / 98

Thinking on, DeepSeek's recommended sampling, mini-SWE-agent with a 100-step limit, graded with the official harness.

Less thinking at every effort setting

DeepSeek-V4.1-Flash has a reasoning-effort dial from 1 to 100 (default 75). So the fair comparison is at matched settings. Accuracy and mean thinking tokens per item, same questions, same serving setup, one run each:

BIG-Bench Hard (540 tasks)

Effort	DeepSeek-V4.1-Flash	AZ-Focus	Thinking tokens
75 (default)	86.9% · 888 tokens	84.3% · 393 tokens	56% less
50	87.8% · 514	86.7% · 238	54% less
35	88.0% · 391	86.1% · 197	50% less
25	86.3% · 298	87.0% · 189	37% less

MMLU-Pro (500 questions, 14 subjects)

Effort	DeepSeek-V4.1-Flash	AZ-Focus	Thinking tokens
75 (default)	87.2% · 1,692 tokens	87.4% · 758 tokens	55% less
50	89.2% · 1,175	86.8% · 527	55% less
35	88.4% · 981	86.2% · 409	58% less
25	88.4% · 699	85.8% · 356	49% less

Single runs on a few hundred questions move by about a point either way. Across the dial AZ-Focus averages about 1.5 points below the original, for roughly half the thinking. If you need every last point on knowledge questions, the original at effort 50 scored highest in our runs (MMLU-Pro 89.2%). If you want the most answers per GPU, AZ-Focus at a low effort setting is the most efficient point we measured.

More benchmarks (default effort)

Benchmark	DeepSeek-V4.1-Flash	AZ-Focus	Thinking tokens
GPQA Diamond (198, 2 runs)	80.8%	86.7%	49% less
HumanEval (164)	99.4%	98.8%
MATH-500 + GSM8K, held-out (450, 3 runs)	96.2%	95.9%	27% less (MATH), 53% less (GSM8K)
AIME 2025 (30, 4 runs)	93.3%	85.0%

Thinking on, DeepSeek's recommended sampling (temperature 1.0, top-p 0.95), served with SGLang. None of these benchmarks was used to build the model. GPQA: the original often ran out of its 16k thinking budget; AZ-Focus rarely does.

Usage

Use it exactly like deepseek-ai/DeepSeek-V4.1-Flash: same serving engines, parallelism, chat template and recommended sampling. The difference shows in thinking mode, and it stacks with the reasoning-effort setting (chat_template_kwargs: {"reasoning_effort": N} in SGLang). With thinking off, output length is unchanged: the edit shortens deliberation, not answers.

Limitations

Hard competition maths is the clear trade-off: AIME 2025 scored about 8 points lower in our runs (4 runs each). If olympiad-style maths is your main use, use the original.
Long multi-hour agent sessions, other languages and very long contexts haven't been measured separately yet.
If you need the model's full self-verification trail, use the original.

License

MIT, as for the original model. Based on DeepSeek-V4.1-Flash by DeepSeek-AI (see LICENSE, DeepSeek_V41_Tech_Report.pdf).

Questions, or want to build on this with us? hello@axonzeta.ai

Downloads last month: 34

Safetensors

Model size

763B params

Tensor type

BF16

F32

F8_E4M3

Model tree for AxonZeta/DeepSeek-V4.1-Flash-AZ-Focus

Base model

deepseek-ai/DeepSeek-V4.1-Flash

Finetuned

(27)

this model