ayjays132/Phillnet-Mini-Max

◈ Phillnet Mini Max

Focused Text + Vision Reasoning with Deep Adaptive Generation

A deliberately lean multimodal model for strong text work and still-image understanding.
Write, reason, explain, transform, code, generate complete HTML, and answer image-grounded questions—without carrying an image or video synthesis stack.

---

◇ The Mini Max proposition

Phillnet Mini Max is the focused text-and-vision edition of Phillnet Mini Omni Max. It retains the language backbone, the visual-understanding route, the native Qwen tokenizer contract, and adaptive reasoning controls while intentionally excluding diffusion and synthesis subsystems. The result is a smaller operational surface for applications that need high-quality language work plus image understanding, not image or video generation.

Best fit: conversational AI, writing assistants, technical explanation, code and single-file HTML generation, text transformation, visual question answering, image-grounded reasoning, and self-hosted multimodal APIs.

Text Intelligence  ·  Still-Image Understanding  ·  Adaptive Reasoning  ·  Production-Ready Local API

---

✦ Capability surface

Capability	Phillnet Mini Max behavior	Practical use
Text generation	Enabled through DendroForCausalLM.generate(...).	Conversation, drafting, rewriting, summaries, structured text, and instruction following.
Reasoned text work	Five selectable effort modes with max as the persisted default.	Short direct answers through long-form planning, technical analysis, and code generation.
Code and HTML	Generates ordinary text tokens, including self-contained code and HTML documents.	Landing pages, dashboards, UI prototypes, scripts, documentation, and configuration templates.
Still-image understanding	Enabled through the local DendroVisionProcessor and retained visual encoder.	Image description, visual Q&A, color and spatial questions, and image-grounded chat.
OpenAI-style service	Included FastAPI endpoint accepts text and optional inline base64 images.	Controlled self-hosted product integration.
Image / video synthesis	Not included. SDXL weights, U-Net, VAE, diffusion schedulers, and synthesis APIs are absent.	Keeps deployment focused on language and visual understanding.
Audio, tools, agents, remote URL fetch	Not exposed by the lean service.	Reduces the deployed attack and dependency surface.

---

✦ Text capabilities, in depth

The language path is first-class in this release. Use it for direct chat, long-form writing, editing, code generation, schema-oriented output, and reasoning-heavy tasks. The model processes the same chat format for text-only and multimodal turns, so an application can begin in text mode and introduce images only when a task requires visual evidence.

Text workflow	Suggested mode	Why it fits
Classification, short formatting, lightweight extraction	direct	Minimizes deliberation for fast, bounded responses.
Summaries, rewrites, concise explanations, ordinary Q&A	low	Adds modest reasoning without the maximum latency profile.
Technical explanations, multi-step planning, nontrivial coding	medium	Balanced answer depth and deliberation.
Architecture reviews, complicated debugging, detailed specifications	high	Allocates deeper internal reasoning.
Long-form HTML, substantial code, complex written deliverables	max	Uses the release default: adaptive private reasoning plus an 8,192-token visible-answer ceiling.

Text-only quick start

pip install -r requirements.txt

import torch
from transformers import AutoModelForCausalLM, AutoProcessor

model_id = "ayjays132/Phillnet-Mini-Max"

processor = AutoProcessor.from_pretrained(
    model_id,
    trust_remote_code=True,
)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
).eval()

messages = [
    {
        "role": "system",
        "content": "You are a precise writing and technical reasoning assistant.",
    },
    {
        "role": "user",
        "content": "Write a concise product brief for a privacy-first research workspace.",
    },
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        reasoning_effort="medium",
        max_new_tokens=700,
        do_sample=False,
        use_cache=True,
    )

prompt_tokens = inputs["input_ids"].shape[1]
answer = processor.tokenizer.decode(
    output[0, prompt_tokens:],
    skip_special_tokens=True,
)
print(answer)

One-shot HTML and code generation

For full-page HTML, use max and allow enough answer tokens for the entire document. The included gallery contains three standalone outputs—two application interfaces and one landing page—created as self-contained HTML with inline CSS and JavaScript.

html_request = [
    {
        "role": "user",
        "content": (
            "Create a complete, responsive, single-file HTML dashboard for a solar-energy "
            "operations team. Use inline CSS and JavaScript. Include a metrics row, a small chart, "
            "status alerts, and a working theme toggle. Return only the HTML document."
        ),
    }
]

inputs = processor.apply_chat_template(
    html_request,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        reasoning_effort="max",
        max_new_tokens=8192,
        do_sample=False,
        use_cache=True,
    )

html = processor.tokenizer.decode(
    output[0, inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)
open("solara-dashboard.html", "w", encoding="utf-8").write(html)

---

◇ Adaptive reasoning

Adaptive reasoning is enabled by default. The model separates internal deliberation from the final answer: its private phase can use the active context budget and end naturally, while the returned answer is bounded independently.

Persisted default	Value	Operational meaning
Default effort	max	The model uses its highest configured effort profile when a caller does not override it.
Adaptive private reasoning	true	Private deliberation can expand within the active context boundary.
Private-reasoning policy	adaptive_context	Reasoning is governed by real context/cache capacity rather than a short hidden caller cap.
Visible answer ceiling	8,192 tokens	Supports substantial documents and code while retaining a predictable returned-output bound.
Active context/cache window	32,768 tokens	Shared physical budget for prompt, private deliberation, and answer allocation.

Use the right budget. For short interactive work, explicitly select direct or low. For whole pages, multi-section documents, or more demanding code, preserve max and permit a correspondingly larger visible-answer budget.

---

◇ Visual understanding

The bundled DendroVisionProcessor is local to this repository and uses the checkpoint’s native Qwen tokenizer contract. It creates visual patches, inserts image placeholders at the correct point in the conversation, and routes visual features into the retained model path.

from PIL import Image

image = Image.open("scene.png").convert("RGB")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {
                "type": "text",
                "text": "Describe the objects in this image and explain their relative positions.",
            },
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        reasoning_effort="medium",
        max_new_tokens=512,
        do_sample=False,
        use_cache=True,
    )

answer = processor.tokenizer.decode(
    output[0, inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)
print(answer)

The visual path is intended for still-image understanding. It does not provide image creation, image editing, video generation, audio understanding, browser tools, or remote image URL fetching.

---

✦ Release gallery

_{Horizon Tasks — Direct  ·  Solara Grid — Medium  ·  Orbit Notes — Adaptive Max}

Example	Focus	Source
Horizon Tasks	Dark productivity landing-page composition.	Open HTML
Solara Grid	Responsive operational dashboard with metrics, CSS chart, alerts, and a theme control.	Open HTML
Orbit Notes	Long-form adaptive one-shot landing page.	Open HTML

---

◇ Verification snapshot

The release was checked at the model, processor, reasoning, visual-input, API, and package levels. The validation evidence is retained in RELEASE_VALIDATION.md, ADAPTIVE_DEFAULTS_REPORT.md, and COHERENCE_REPAIR_REPORT.md.

Area	Verified result
Core checkpoint	DendroForCausalLM loads in BF16.
Native tokenizer	Local Qwen tokenizer matches the checkpoint’s 248,320-token vocabulary.
Text coherence	Deterministic text probes returned Paris, 4, and blue.
Local processor	AutoProcessor resolves to DendroVisionProcessor.
Image grounding	Controlled solid-color probes produced the corresponding colors.
Spatial grounding	A red-left / blue-right probe returned red for left and blue for right.
Long text / HTML	A complete 2,991-token Orbit Notes HTML page was produced after 2,944 private-reasoning tokens.
Lean scope	SDXL weights, code, dependencies, and synthesis APIs are absent.

Open the retained-checkpoint integrity record

model.safetensors
SHA-256: f1a913f99f8ce921c1aa79982a09eaee6744448766f09a4756751e2d4b9342fc
Size:    1,763,655,304 bytes

The repository’s CHECKSUMS.sha256 and RELEASE_MANIFEST.json provide the corresponding reproducibility records.

---

◇ Architecture and package layout

PHILLNET MINI MAX — TEXT + VISION EDITION
│
├── Language path       Dendro causal language model
├── Vision path         Retained visual encoder + local DendroVisionProcessor
├── Token vocabulary    248,320 native Qwen-compatible tokens
├── Reasoning policy    Adaptive private deliberation; max by default
├── Answer policy       Up to 8,192 visible answer tokens
├── Checkpoint          1.76 GB SafeTensors, BF16-capable load path
└── Excluded systems    SDXL, U-Net, VAE, diffusion, image/video synthesis,
                        audio, agents, tools, and remote image fetching

Repository path	Purpose
model.safetensors	Retained language and visual-understanding checkpoint.
config.json	Model configuration, adaptive defaults, and custom auto mappings.
processing_dendro_omni.py	Local native-tokenizer text-and-image processor.
modeling_dendro_omni.py	Custom language and multimodal generation runtime.
server.py	Text-and-vision-only OpenAI-style FastAPI service.
examples/	Complete HTML examples and visual previews.
PRODUCTION.md	Deployment topology, security controls, and operations guide.
HF_UPLOAD.md	Hugging Face upload and verification workflow.
RELEASE_MANIFEST.json	Release identity, artifact hashes, capability boundary, and defaults.

---

◇ Local API and production deployment

The included service exposes only POST /v1/chat/completions, with text and optional inline base64 still images.

pip install -r requirements.txt -r requirements-server.txt
uvicorn server:app --host 127.0.0.1 --port 8000

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  --data '{
    "messages": [
      {
        "role": "user",
        "content": "Draft a concise engineering handoff for a feature-flag rollout."
      }
    ],
    "reasoning_effort": "medium",
    "max_tokens": 700
  }'

For deployment, the repository includes a non-root Dockerfile, loopback-bound docker-compose.yml, readiness checks, optional API-key protection, explicit CORS configuration, image/request limits, and a single-generation lock. Start with:

cp .env.example .env
# Set PHILLNET_API_KEY and PHILLNET_CORS_ORIGINS before public exposure.
docker compose up --build -d

Read PRODUCTION.md before placing the service behind a public endpoint.

---

◇ License and provenance

This derivative retains the upstream Apache-2.0 designation. It is based on the Phillnet Mini Omni Max release and preserves only its text and still-image-understanding routes. See the upstream model card for the source release context. 1

Attribute	Value
License	Apache-2.0
Upstream reference	ayjays132/Phillnet-Mini-Omni-Max
Release repository	ayjays132/Phillnet-Mini-Max
Loading requirement	trust_remote_code=True