language:
- ru
- en
license: apache-2.0
base_model: heretic-org/Qwen3.8-27B-heretic-ara
pipeline_tag: image-text-to-text
inference: false
tags: - sovereign
- splash
- apple-silicon
- qwen3_5
- 4-bit
- vision
- tool-use
- dflash2

Sovereign Uncensored ARA 27B Splash Q4 — v2.0.1-rc2
A local 27B text and vision model for Apple Silicon, with an optional desktop interface and agent tools.
Sovereign combines a Q4 package of the Heretic ARA 27B model with Splash for local inference. Sovereign Desktop adds Hermes and Companion for conversations, coding, documents, memory and research. This RC2 package runs on your Mac when selected through the local Splash provider. Desktop can also use separately configured remote models; network tools connect to their respective external services.
V2 RC2 · ARA 27B Q4 · Splash 1.3.0 + DFlash2 · Companion 0.3.5.4 RC14 · Apple Silicon
Run the model directly with Splash, or use the separately distributed Sovereign Desktop for Hermes and Companion. The Desktop installer is designed to connect an existing complete model package without downloading the weights again.
Measured locally on Apple M4 Max / 48 GB: median ~50 tok/s · up to 106 tok/s across 38 live engine-decode samples (October 23, 2026). Measurement scope and current vision timings.
Key features
- Text and images. Generate and edit text, and ask questions about images.
- Code and documents. With Sovereign Desktop, read selected project files, inspect and check code, and search local materials.
- Profiles and tool controls. Desktop's four built-in profiles cover general work, coding, research and documents. New and cloned profiles connect to Companion automatically; explicitly disabled connections are respected. Tool sets can be enabled manually, with automatic selection limited to enabled sets.
- Persistent memory and saved tasks. Companion can store useful facts separately from task progress and resume saved work from checkpoints.
- Local inference. Splash runs the compatible model on the Mac. Hermes displays a live decoding speed in tokens per second, and compatible requests can reuse cached context.
- Updates and stored data. SOUL, settings and history are stored separately from the application bundle. Companion activation and rollback were checked in an isolated installation. The separate Desktop RC14 installer passed an RC12 → RC14 → RC12 installation/receipt cycle in a macOS VM; full first-run acceptance on a clean physical Mac remains outstanding.
Companion tools
| Tool | Function |
|---|---|
| Compact Repair | Applies targeted Python edits, checks the source version and validates the result, then saves a new copy. |
| Serena | Finds files, functions and references across a project, and supports targeted code edits. |
| Exa | Searches the web for information and code examples, and retrieves page content. |
| Code Guard | Checks Python syntax, imports in an isolated environment, and specified tests. |
| Context7 | Retrieves documentation and usage examples for libraries and frameworks. |
| Playwright | Opens and reads web pages, clicks elements and fills forms. |
| Hugging Face | Searches for models, datasets, research papers and documentation. |
Memory is available as a separate tool set. Network tools may require credentials and transmit requests to their providers.
Downloads and demos
The model files are available directly in this public repository. Follow the standalone setup below to run RC2 with Splash.
- Companion 0.3.5.4 RC14 source ZIP — optional developer distribution for an existing compatible Hermes host. Read COMPANION.md for prerequisites, installation and rollback.
- Release checksums and verification report.
Code tools: Compact Repair + Code Guard
A checkout calculation returns the discount instead of the final price. Sovereign proposes one targeted edit, preserves the original file, and checks the candidate against the same two assertions. Both pass.
Download the code tools demo · English subtitles · Русские субтитры
More Companion demonstrations
- Serena: locate a function and verify its callers.
- Task Continuation: pause an inventory reconciliation and complete it in a new chat.
- Exa + Playwright: find and inspect the official Git manual.
These explainers use animated diagrams and enlarged evidence from Sovereign Desktop 0.3.5.4 RC9 recordings. English narration was generated with ElevenLabs through fal.ai; English captions use observed Scribe V2 word timings, with optional English/Russian subtitle tracks. These are not new RC14 installer recordings. See MEDIA-PRODUCTION.md.
RC2 vision: review a release dashboard
The local model reads an engineering dashboard containing a dependency map, a metrics table, a latency chart and release gates. It traces required and optional dependencies, reads the candidate metrics and applies the displayed rules. All 10 checked fields matched on each of two images: candidate r42 at 310 ms is correctly rejected against a 250 ms limit; candidate r43 at 220 ms is correctly accepted while Search remains optional. See the inputs, raw API answers and independent checks below. This is a narrow authored demonstration, not a general accuracy benchmark.
Download the engineering vision demo · Test images and results · English subtitles · Русские субтитры
Components
Model: a 27B Qwen-family model with Heretic ARA modifications, packaged in affine Q4 for Splash. Vision provides image input; DFlash2 supplies a draft model for speculative decoding.
Splash 1.3.0: the pinned local inference engine for this model package.
Hermes: the conversation interface, history and agent loop, allowing the model to use enabled tools and their results.
Companion: profiles, tool controls, local documents, memory and saved tasks.
Sovereign Desktop: the macOS installer and launcher for Hermes and Companion. Core utilities are bundled; additional dependencies are prepared on first launch.
Run the model without Hermes or Companion
You can run the model with Splash 1.3.0 alone and use its browser chat or local OpenAI-compatible API. Hermes, Companion, LM Studio and Homebrew are not required. Companion profiles, memory, saved tasks and tools are separate Desktop features; tool calls need a client to execute them.
You need an Apple Silicon Mac (M3 or newer), macOS 26.4 or later, and the complete RC2 package. The model files occupy approximately 16.19 GiB. The configuration below was tested on an M4 Max with 48 GB of unified memory and macOS 26.6.2. A minimum memory requirement for this package has not been established.
Install the compatible engine
Download the official 1.3.0 archive and verify its checksum:
mkdir -p "$HOME/Sovereign-standalone"
cd "$HOME/Sovereign-standalone"
curl --fail --location --output splash-1.3.0-arm64-macos26.tar.gz \
https://github.com/incoai/splash/releases/download/1.3.0/splash-1.3.0-arm64-macos26.tar.gz
printf '%s %s\n' \
9cbf463598367c912ee17eaba27cba4be3d324d688541f7ed5106ed9b2bd691d \
splash-1.3.0-arm64-macos26.tar.gz | shasum -a 256 -c && \
tar -xzf splash-1.3.0-arm64-macos26.tar.gz
Continue only if the checksum says OK. Use this pinned version: Splash 1.3.1 removed support for Splash-packed packages.
Download or select the model
To download this model from the Hub, run the following:
This repository is public. Download the complete RC2 model without signing in:
cd "$HOME/Sovereign-standalone/splash-1.3.0-arm64-macos26"
HF_REPO_ID="proofofvitalik/Sovereign-Uncensored-ARA-27B-Splash-Q4-v2.0.1-rc2"
MODEL_DIR="$HOME/Sovereign-standalone/model-rc2"
./python/bin/python3 -B -m huggingface_hub.cli.hf download "$HF_REPO_ID" --local-dir "$MODEL_DIR"
If you already have the complete RC2 package, skip the download and set MODEL_DIR to the folder that directly contains manifest.json, replacing the example path:
MODEL_DIR="$HOME/path/to/Sovereign-Uncensored-ARA-27B-Splash-Q4-v2.0.1-rc2"
Check the manifest:
printf '%s %s\n' \
fbf119258b2f34e66d7dca112f14ac2d73614cdc3049d4a4710aec0069d20606 \
"$MODEL_DIR/manifest.json" | shasum -a 256 -c
Continue only if the checksum says OK. This checks the manifest, not each weight file.
Start the model
In the same Terminal window:
cd "$HOME/Sovereign-standalone/splash-1.3.0-arm64-macos26"
./python/bin/python3 -B -m server.server "$MODEL_DIR" \
--tokenizer "$MODEL_DIR/tokenizer" \
--binary "$PWD/engine/splash" \
--model local/sovereign-rc2 \
--host 127.0.0.1 --port 18019 \
--max-context 4096 --max-memory 20G \
--default-reasoning-effort none --idle-release 1m \
--request-timeout 120 --queue-size 1
When the server prints Ready, open http://127.0.0.1:18019. Keep Terminal open while using the model; press Ctrl+C there to stop it. Splash's notice that PyTorch is unavailable does not prevent this inference path from working. If the port is in use, choose another free port and update the URL and API command accordingly. The example admits one request at a time; concurrent requests can receive HTTP 503. Increase --queue-size if your client needs concurrency and the Mac has sufficient resources.
To send a request from another Terminal window:
curl --fail --silent --show-error http://127.0.0.1:18019/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model":"local/sovereign-rc2","messages":[{"role":"user","content":"Reply with exactly the word READY."}],"max_tokens":24,"temperature":0}'
OpenAI-compatible clients can use base URL http://127.0.0.1:18019/v1 and model ID local/sovereign-rc2. See STANDALONE.md for detailed setup and verification notes.
Hermes, Companion and Sovereign Desktop
The Companion ZIP in this repository contains source, pinned dependency locks, installer, rollback utility and third-party attribution. It requires an existing compatible Hermes host; it does not install Hermes, Splash or model weights. It was validated with the Sovereign Hermes core at d923a5eb487a17c0a68801bba9ef0f56c81040a8. Compatibility with arbitrary upstream Hermes releases is not established. See COMPANION.md.
Sovereign Desktop is the separate macOS application distribution. Its installer is not included in this model-and-Companion release. Desktop's application integration, managed local provider and cache hooks require the matching host. The Desktop source repository is currently private. The standalone Splash instructions on this page provide the complete public path to running the model.
Measured speed
Apple M4 Max · 48 GB unified memory · Sovereign RC2 Q4 + DFlash2
| Saved measurement | Observed generation rate | Scope |
|---|---|---|
| Text-generation check · October 23, 2026 | 35–106 tok/s, median ~50 | 38 live native engine-decode samples from an earlier Desktop 0.3.5.4 build |
| Engineering vision test · November 16, 2026 | 39.3–46.5 tok/s | 2 local image requests through official Splash 1.3.0 |
The text counter measures whole-engine decoding and excludes queueing, prompt processing and tool time. The vision row is a per-request output rate after the first token batch; its initial wait was 19.0–22.5 seconds. These are different workloads and metric scopes, not an A/B comparison or a guarantee of total response time. See SPEED.md for settings, methodology and all numeric evidence.
Measurements and verification
On November 16, 2026, a new local RC2 image test reviewed a synthetic engineering dashboard with dependencies, service metrics, a chart and release rules. It returned the correct structured values and decisions for both independent images: 10/10 checked fields in each case. The candidate and latency were changed between the images; the prompt stayed the same. See VISION-AUDIT.md for the exact scope, original inputs, raw responses and checks.
On November 13, 2026, the official Splash 1.3.0 archive loaded RC2 on an Apple M4 Max / 48 GB Mac. Text returned READY, two image inputs returned the expected red/blue colors, and a generated add_numbers(137, 205) call was executed locally and returned to the model, which answered 342. The browser chat endpoint also responded. These are smoke checks, not a general capability benchmark. All 84 target/draft/vision/tokenizer/manifest files (17,381,703,602 bytes) matched the package inventory; large-file hashes also matched Hugging Face metadata. Companion's 88 inventory entries matched its source ZIP; 127 Python tests and the browser-policy, Desktop controller, metrics, task-evidence and task-store process checks passed. See RELEASE-QA.md.
On October 30, 2026, the official standalone Splash 1.3.0 archive loaded RC2 on an M4 Max / 48 GB Mac and returned READY through its local API. A separate API request generated a structured tool call; the test did not execute an external tool. The server also delivered its browser chat page. In a separate image-input smoke test, the downloaded RC2 correctly identified red and blue circles in two synthetic images. On October 31, 2026, isolated Hermes/Companion checks from the Desktop candidate also passed image analysis, scoped file reading, missing-source handling and saved-task continuation after restarting the test backend. The subsequent RC13 candidate passed real model/tool calls, the metrics API and Companion rollback/reactivation in a separate test profile on the existing Mac.
Provenance
Sovereign integrates and adapts existing components; no additional model training was performed as part of this project.
- Heretic ARA / Qwen3.8-27B — target model, tokenizer and vision components.
- Qwen3.8-27B-DFlash2 — draft model for speculative decoding.
- Hermes Desktop / Agent — the basis of the interface and agent loop, with Sovereign modifications.
The upstream name contains Qwen3.8; the architecture is identified as qwen3_5 in the configuration. The Uncensored label refers to the model's upstream lineage.
Licensing
The Apache-2.0 metadata on this card applies to the model package, which retains its LICENSE and NOTICE.md. Original Sovereign installer and Companion additions are also licensed under Apache-2.0. The desktop distribution combines components under their own licenses:
| Component | License |
|---|---|
| Sovereign model package | Apache-2.0; upstream attribution retained in LICENSE and NOTICE.md. |
| Original Sovereign installer and Companion additions | Apache-2.0. |
| Hermes and Hermes-derived desktop/CLI files | MIT. |
| Electron | MIT, with separate Chromium and third-party notices. |
| Python runtimes | Python/CNRI terms and included libraries' licenses. |
| Node.js / npm | Node.js combined license text; npm uses Artistic-2.0, with separate dependency licenses. |
| uv / ripgrep | MIT OR Apache-2.0 for uv; MIT OR Unlicense for ripgrep. |
| Git | GPL-2.0. |
| Included FFmpeg build | GPL-3.0-or-later, according to its build configuration. |
| Electron's separate shared FFmpeg library | LGPL-2.1-or-later; corresponding source and build records are covered separately from the command-line FFmpeg build. |
The Desktop distribution includes third-party notices and license texts. Splash, MCP packages, browser dependencies and other components retain their respective licenses and terms. External services have their own usage terms. The model's Apache-2.0 label does not relicense the full macOS distribution.
Technical specification and checksums
| Parameter | Value |
|---|---|
| Model | 27B, affine Q4, group size 64 |
| Format | splash-packed-q4, manifest schema 3 |
| Vision | BF16 |
| Draft | DFlash2, Q4/group64 |
| Engine | Splash 1.3.0, pinned version |
| Model version | v2.0.1-rc2 |
| Target model revision | 2dc9b364104881cbb85e390f00195ba6b9d745e9 |
| DFlash2 revision | 015e795645c74b1a0eeef3b570031fb62e769bc5 |
The packed model is intended for the specified Splash version. It cannot be loaded with a standard MLX loader or transformers.from_pretrained(); another engine requires a separate compatibility check. A ready-to-use Hugging Face Inference API provider is not claimed.
The main text weights were not requantized during packaging. Vision tensors were converted through name and layout transformations; the draft model was packed in Q4. The chat template was adapted for thinking disabled by default and tool use. Upstream training data composition has not been independently verified.
SHA-256 of model manifest.json: fbf119258b2f34e66d7dca112f14ac2d73614cdc3049d4a4710aec0069d20606.
Identifier for the set of 81 model runtime artifacts: 224581e131cc3e520c9c3174d23d3a015c174b466a6e55fcc4428f1eef09342e.