OsaurusAI/Raptor-V1-26B-A4B-VL-JANG

Raptor V1 · 26B-A4B · JANG

Mixture-of-experts · 26B total / ~4B active · Apple Silicon · mixed-precision JANG · ~13 GB on disk

---

What this is

A JANG-quantized build of Gemma 4 26B-A4B for local inference on Apple Silicon through Osaurus. It is packaged for the Osaurus tool surface — files, shell, sandbox, database, browser, AppleScript and computer use — and runs entirely on-device.

Mixture-of-experts: a fraction of the total parameters is active per token, so decode cost tracks the active set rather than the full model. That is what lets a 26B-class model sit alongside your actual work on a laptop instead of monopolising it.

Quantization

method	JANG affine (mx.quantize backend)
attention / router	8-bit
routed experts gate_proj	4-bit
routed experts up_proj / down_proj	2-bit (4-bit on the first two and last two layers)
group size	64
per-module overrides	180 (326 quantized modules total)
high-precision passthrough	624 tensors
embeddings	8-bit affine (tied to the output head)
norms	gemma4_scale_shift_zero convention
vision embedder	preserved, fp16 passthrough (early fusion)
on disk	~13 GB, 13 shards
context	262,144

Mixed precision rather than a single global bit width: the modules where error changes behaviour rather than degrading it smoothly — norms, embeddings, the vision embedder — are kept in high precision, and the per-module override map carries the rest.

The first of a mixture of models

Raptor One is the first release in what will become a mixture of models — a set of specialists Osaurus routes between, rather than one model asked to be everything. A coder is planned, and a dedicated VL/OCR model.

Raptor One already carries vision, but reading a dense scanned document well is a different job from driving an agent loop, and being built for one does not make a model good at the other. Raptor One's role in the set is orchestrator: hold the session, plan the work, call the tools, recover when a call fails, and hand the specialist jobs to specialists once they exist.

That is why it is built around agentic behaviour and error recovery rather than benchmark breadth — fast and disciplined ahead of encyclopaedic.

Requirements

• Apple Silicon (M-series)
• macOS with Osaurus, or any MLX-based runtime
• ~13 GB free disk, and enough free memory to hold the resident set alongside your other work

Sampling

Defaults ship in generation_config.json and jang_config.json. Gemma-4 is sensitive to them — in particular top_k, without which output quality degrades noticeably. Use the shipped values unless you have measured a reason not to.

Usage

In Osaurus, place the bundle under your models directory and select it from the model picker. For agentic work, bind a working folder from the chat rail — the file and shell tools are injected per-folder, so without one the model has no filesystem surface to act on.

Licence and attribution

Apache-2.0, inherited from the base model. Google publishes Gemma 4 under Apache-2.0 — see google/gemma-4-26B-A4B-it, whose model card declares license: apache-2.0. This build carries the same licence; nothing about modifying it changes the terms.

Separately from the licence, Google's Gemma usage terms set out a prohibited-use policy that applies to Gemma models and their derivatives. Using this build means honouring that policy as well.

Base model © Google DeepMind. Gemma is a trademark of Google LLC. Quantization, packaging and the Osaurus integration are ours; the weights are Google's work.

Benchmarks

Not yet published for this build. Numbers will be added here once measured on this artifact — figures from other models or other quantizations are not transferable and are deliberately omitted rather than estimated.