Mira-1-XL API & Inference Endpoint

What’s inside

Architecture: Qwen3.5 multimodal (Qwen3_5ForConditionalGeneration)
Precision: BF16 weights
Modalities: image + text → text

Merge recipe (DARE-TIES)

This model is a 3-way DARE-TIES merge (drop + rescale, then weighted delta sum), with per-source sparsity (“density”) and weights.

Important: DARE-style merges are stochastic unless you fix RNG seeds. Your merge may differ slightly if you re-run.

Base

Qwen/Qwen3.5-27B

Merge sources + parameters

Table with columns: Source model, Role / flavor, density, weight
Source model	Role / flavor	density	weight
`Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled`	reasoning + structured solutions	0.70	0.40
`TeichAI/Qwen3.5-27B-Claude-Opus-4.6-Distill`	instruction following + general ability	0.65	0.35
`TeichAI/Qwen3.8-27B-Fable-Distill`	creative writing / planning style	0.60	0.30

Normalization: enabled (per-tensor norm stabilization)
Output dtype: bf16

Intended use

Great for:

multimodal Q&A (“what’s in this image?”, “summarize this screenshot”, etc.)
long-form writing (stories, scripts, outlines)
agent-style planning (multi-step task decomposition)
general chat + instruction following
“reasoning-ish” tasks where base Qwen can be too shallow

Not ideal for:

safety-critical domains (medical/legal/financial advice) without human review
guaranteed factuality (hallucinations still happen)
strict reproducibility (DARE merges are not deterministic unless seeded)

Usage (Transformers)

1) Quick pipeline (image-text-to-text)

python
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="Bc-AI/mira-1-xl")  # <- change to your repo id

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? Answer in one sentence."}
    ]
  }
]

out = pipe(text=messages, max_new_tokens=128)
print(out)

What’s inside

Architecture: Qwen3.5 multimodal (Qwen3_5ForConditionalGeneration)
Precision: BF16 weights
Modalities: image + text → text

Merge recipe (DARE-TIES)

This model is a 3-way DARE-TIES merge (drop + rescale, then weighted delta sum), with per-source sparsity (“density”) and weights.

Important: DARE-style merges are stochastic unless you fix RNG seeds. Your merge may differ slightly if you re-run.

Base

Qwen/Qwen3.5-27B

Merge sources + parameters

Table with columns: Source model, Role / flavor, density, weight
Source model	Role / flavor	density	weight
`Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled`	reasoning + structured solutions	0.70	0.40
`TeichAI/Qwen3.5-27B-Claude-Opus-4.6-Distill`	instruction following + general ability	0.65	0.35
`TeichAI/Qwen3.8-27B-Fable-Distill`	creative writing / planning style	0.60	0.30

Normalization: enabled (per-tensor norm stabilization)
Output dtype: bf16

Intended use

Great for:

multimodal Q&A (“what’s in this image?”, “summarize this screenshot”, etc.)
long-form writing (stories, scripts, outlines)
agent-style planning (multi-step task decomposition)
general chat + instruction following
“reasoning-ish” tasks where base Qwen can be too shallow

Not ideal for:

safety-critical domains (medical/legal/financial advice) without human review
guaranteed factuality (hallucinations still happen)
strict reproducibility (DARE merges are not deterministic unless seeded)

Usage (Transformers)

1) Quick pipeline (image-text-to-text)

python
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="Bc-AI/mira-1-xl")  # <- change to your repo id

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? Answer in one sentence."}
    ]
  }
]

out = pipe(text=messages, max_new_tokens=128)
print(out)

Mira-1-XL

README

What’s inside

Merge recipe (DARE-TIES)

Base

Merge sources + parameters

Intended use

Usage (Transformers)

1) Quick pipeline (image-text-to-text)

Explore FriendliAI today

README

What’s inside

Merge recipe (DARE-TIES)

Base

Merge sources + parameters

Intended use

Usage (Transformers)

1) Quick pipeline (image-text-to-text)