New · live on Hugging Face

FR-Blaze-9B

A Thin Language Model for growth. A Gemma 2 9B lens for SEO and paid-platform marketing: audits, Search Console and Google Ads framing, on-page and backlink analysis, RAG-grounded, with deterministic guardrails that refuse black-hat tactics by design.

TLM · Growth & SEO 9B · Gemma 2 base Gemma license
StatusShipped
Parameters9B · 4-bit
BaseGemma 2 9B
Eval83.6% (lens)
Parameters
9B
Quantization
4-bit
Base Model
Gemma 2
Lens Eval
83.6%
License
Gemma
The Lens Pattern

FR-Blaze as a growth lens

FR-Blaze sits in front of your SEO and paid-marketing workflow — scoping, auditing, and routing queries before they reach an expensive general-purpose model or API.

User query
SEO / Ads question
The Lens
FR-Blaze 9B
LOCAL · ON-DEVICE
If needed
Frontier LLM
SCOPED + ROUTED

Guardrails that refuse black-hat by design

FR-Blaze has deterministic guardrails built into its weights. Black-hat SEO tactics — cloaking, keyword stuffing, link schemes, hidden text — are refused at the model level, not filtered after the fact. No prompt can override this.

Capabilities

Built for growth intelligence

FR-Blaze covers the reasoning tasks that matter in SEO and paid-platform marketing — grounded in real data, without cloud egress.

001

SEO Audits

Analyse on-page factors, surface technical issues, flag thin content, and prioritise fixes ranked by estimated impact — grounded in your actual site data.

002

Search Console Intelligence

Interpret GSC exports to surface CTR opportunities, impression-rank gaps, and cannibalisation patterns with structured recommendations.

003

Google Ads Framing

Frame campaign briefs, audit ad copy alignment with landing pages, and flag Quality Score risk factors — all via the lens before the expensive API call.

004

Backlink Analysis

Classify link profiles by authority and risk, flag toxic patterns, and surface anchor-text distribution anomalies from structured export data.

005

RAG-grounded Answers

Pair FR-Blaze with a retrieval layer over your own SEO data, historical reports, or knowledge base — answers stay grounded in your sources, not hallucinated.

006

Black-hat Refusal

Cloaking, keyword stuffing, link schemes, hidden text, doorway pages — refused at the weight level. Deterministic. Not a post-hoc filter.

Architecture

How FR-Blaze works

Distilled from frontier SEO and growth reasoning into a Gemma 2 9B base, running locally with no cloud dependency.

01

Growth Domain Corpus

Trained on a proprietary corpus of SEO documentation, Search Console exports, Google Ads policy, backlink datasets, on-page audits, and growth case studies — curated and reviewed in-house by SEO and paid-media specialists.

02

Gemma 2 9B Base

Built on Gemma 2 9B — a capable open base model from Google DeepMind. The larger parameter count relative to our 1.7B models gives FR-Blaze stronger multi-step reasoning for complex SEO and paid-media analysis tasks.

03

Frontier Distillation + LoRA

Frontier models with strong marketing and SEO reasoning serve as teachers. LoRA fine-tuning on Apple Silicon via MLX locks in domain judgment — jurisdiction of intent, keyword classification, bid strategy framing — not just surface vocabulary.

04

Deterministic Guardrails

Black-hat refusal is built into the fine-tuning targets, not a runtime filter. The model was trained to refuse these outputs unconditionally — no system prompt modification can override it.

05

Standalone or Lens Deployment

Run FR-Blaze as a fully on-device model for growth queries, or wire it as a lens in front of a frontier model API. In Lens mode it scopes intent, classifies query type, and routes — cutting frontier calls to only what genuinely needs them.

Benchmarks

83.6% on the lens eval

Evaluated on SEO and growth-domain reasoning tasks. FR-Blaze outperforms general-purpose models significantly larger on its target domain.

Task FR-Blaze 9B GPT-4o-mini Llama 3.1 8B
SEO intent classification 91.2% 82.4% 74.1%
On-page issue detection (F1) 88.4 79.7 68.3
GSC opportunity surfacing 86.9% 77.3% 65.8%
Black-hat refusal rate 99.8% 71.2% 64.5%
Lens eval (overall) 83.6% 74.1% 61.9%
Hardware Requirements

Runs on what you already have

Apple Silicon

Recommended · MLX runtime
ChipM2 or later (M3 recommended)
RAM16 GB unified (32 GB recommended)
Storage~5.5 GB (4-bit)
Tokens/sec~18 tok/s on M3 Max

CUDA GPU

Supported · llama.cpp / GGUF
GPU VRAM8 GB minimum (16 GB recommended)
RAM16 GB system RAM
Storage~5.5 GB (Q4_K_M)
Tokens/sec~28 tok/s on RTX 4080

Deploy FR-Blaze in
your growth stack.

Download the weights on Hugging Face, run it on-device, or wire it as a lens in front of your existing marketing toolchain.