Blossom-V7.1-35B-A3B

Chat | GitHub

Blossom-V7.1 is a family of open-weight, general-purpose multimodal models designed for local deployment. It combines efficient thinking with two modes (medium and max), tool use with interleaved thinking, and image understanding. It covers everyday conversation, world knowledge, mathematics and reasoning, coding, web development, and data visualization, and supports agentic workflows through tool use.

Key Features

• Efficient Two-Mode Thinking: Supports always-on medium and max thinking, defaulting to max for more thorough reasoning. The medium mode scales reasoning depth to task difficulty, delivering high-quality results with reasoning traces about one-quarter as long as those of Qwen3.5 and one-fifth as long as those of Qwen3.6.
• Tool Use with Interleaved Thinking: Reasons and makes decisions before every tool call, enabling strong performance in agentic tasks.
• Image Understanding: Understands image inputs alongside text.
• Long Context: Handles up to 262,144 tokens (256K); use 131,072 tokens (128K) for best results.
• Faster Inference with MTP: Supports speculative decoding via Multi-Token Prediction (MTP) in both vLLM and llama.cpp.

Important: Blossom-V7.1 uses a custom chat template that differs from the native templates of its Qwen base models. Always use the bundled chat_template; do not replace it with a native Qwen chat template or combine the two templates.

Model Variants

Model	Resources	Base Model
Blossom-V7.1-27B	Demo GGUF	Qwen3.8-27B
Blossom-V7.1-35B-A3B	Demo GGUF	Qwen3.6-35B-A3B
Blossom-V7.1-9B	Demo GGUF	Qwen3.5-9B

Select a variant based on your quality target, inference hardware, and memory budget:

• 27B: The most capable dense option, intended for GPU deployments where the model's weights fit entirely in GPU memory.
• 35B-A3B: The throughput-oriented option for CPU or hybrid CPU/GPU inference, offering a practical balance of quality and speed even with partial CPU offload.
• 9B: The lowest-resource option for memory-constrained GPUs, mobile devices, and lighter workloads.

Post-Training

Blossom-V7.1 is post-trained for general assistant use across everyday conversation, world knowledge, mathematics and reasoning, coding, web development, and data visualization.

The data pipeline uses BlossomData, our open-source framework for flexible, scalable data processing and synthesis. Data is filtered with LLM-as-Judge review and, when applicable, Agent-as-Judge verification. Agent-as-Judge runs in AgentBox, our open-source work environment for AI agents, and uses search, browser interaction, screenshot capture, and code execution as needed. Samples involving web development and data visualization receive additional screening for functionality, usability, and visual quality.

The training data will be released as open source in a future update.

Multi-turn & Reasoning Replay

For multi-turn conversations, always replay the assistant's reasoning together with its answer. Omitting prior reasoning can significantly degrade model performance in subsequent turns.

After initializing the OpenAI-compatible client shown in either server example below, append the complete assistant message rather than rebuilding it from role and content:

messages = [{"role": "user", "content": "Explain why the sky is blue."}]

response = client.chat.completions.create(model="blossom-v7.1", messages=messages)
messages.append(response.choices[0].message.model_dump(exclude_none=True))
messages.append({"role": "user", "content": "Now explain it with an analogy."})

response = client.chat.completions.create(model="blossom-v7.1", messages=messages)

This preserves vLLM's reasoning field, llama.cpp's reasoning_content field, and any tool calls. Configure agent frameworks to retain the complete assistant message in history.

Usage

The examples below use Blossom-V7.1-27B. To switch variants, set MODEL_ID to the corresponding Safetensors repository for Transformers or vLLM, or to the GGUF repository for llama.cpp.

Thinking Modes

Thinking is always on, with two modes: medium and max. The default is max. The bundled chat template maps reasoning_effort values to these modes as follows:

reasoning_effort	Thinking Mode
Omitted (default)	max
low, medium, adaptive (template alias)	medium
All other values	max

Recommended Sampling

The recommended settings are temperature=1.0, top_p=0.95, top_k=50, and repetition_penalty=1.0. The first three are included in generation_config.json and the GGUF metadata, while all three runtimes default to repetition_penalty=1.0. In most cases, leave them unset.

Transformers

Install PyTorch for your hardware, then install:

pip install -U "transformers>=5.12.1" accelerate

This text-only path skips the vision encoder:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "Azure99/Blossom-V7.1-27B"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Explain why the sky is blue in simple terms."}
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    generated_ids = model.generate(
        **inputs,
        max_new_tokens=2048,
    )

generated_ids = generated_ids[:, inputs["input_ids"].shape[1]:]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

vLLM

Install vLLM and the OpenAI client:

pip install -U "vllm>=0.26.0" openai

Start an OpenAI-compatible server:

MODEL_ID=Azure99/Blossom-V7.1-27B

vllm serve "$MODEL_ID" \
  --served-model-name blossom-v7.1 \
  --max-model-len 131072 \
  --gpu-memory-utilization 0.95 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --enable-prefix-caching \
  --speculative-config '{"method":"mtp","num_speculative_tokens":1}'

The tokenizer includes the full chat template for multimodal input, reasoning, and tool calls, and vLLM loads generation_config.json by default. Add --tensor-parallel-size N for multi-GPU serving. Set --max-model-len 262144 if memory allows. MTP is optional; remove --speculative-config to disable it.

Call the server with an image URL using the OpenAI client. vLLM returns parsed reasoning in message.reasoning:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
IMAGE_URL = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"

response = client.chat.completions.create(
    model="blossom-v7.1",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {"url": IMAGE_URL},
                },
                {"type": "text", "text": "Describe this image briefly."},
            ],
        }
    ],
    max_completion_tokens=512,
)

message = response.choices[0].message
print("Reasoning:", getattr(message, "reasoning", None))
print("Answer:", message.content)

To select medium mode with vLLM, add reasoning_effort="medium" to the client.chat.completions.create(...) call above.

llama.cpp (GGUF)

For GGUF inference, use the matching GGUF repository with a current llama.cpp build. Use the embedded chat template; do not pass --chat-template or --chat-template-file. Keep --reasoning on and --reasoning-format deepseek enabled as shown below.

Ollama is not recommended because its current chat implementation does not correctly preserve Blossom's template behavior.

MODEL_ID=Azure99/Blossom-V7.1-27B-GGUF

llama-server \
  -hf "${MODEL_ID}:Q4_K_M" \
  --alias blossom-v7.1 \
  --ctx-size 131072 \
  --parallel 1 \
  --flash-attn on \
  --spec-type draft-mtp \
  --spec-draft-n-max 1 \
  --reasoning on \
  --reasoning-format deepseek \
  --min-p 0

-hf loads the Q4_K_M model and its embedded chat template, and automatically downloads a multimodal projector from the same repository. --min-p 0 disables llama.cpp's default min-p sampler. Set --ctx-size 262144 if memory allows. MTP is optional; remove --spec-type and --spec-draft-n-max to disable it.

With the parsing flags above, llama.cpp returns reasoning in message.reasoning_content and tool calls in message.tool_calls.

Call the server with the OpenAI client:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="no-key")
response = client.chat.completions.create(
    model="blossom-v7.1",
    messages=[{"role": "user", "content": "Find an elegant proof that there are infinitely many primes."}],
    max_tokens=2048,
)

message = response.choices[0].message
print("Reasoning:", getattr(message, "reasoning_content", None))
print("Answer:", message.content)

License

Blossom-V7.1 is released under the Apache License 2.0.