FR-Blaze-9B

A Thin Language Model for online marketing. Grounded, specialized, and running on your laptop.

GROUNDED.   SPECIALIZED.   LOCAL.

---

Contents

1. Overview
2. How it works
3. Specifications
4. Quickstart
5. Intended use
6. Evaluation
7. Limitations
8. Training
9. Citation

Overview

FR-Blaze-9B is a Thin Language Model (TLM) for online and digital marketing, by Fahrenheit Research. Sibling to FR-Forge (manufacturing) and FR-Lex (legal). It is a 4-bit, MLX model: the Gemma 2 9B base with LoRA adapters fused in, designed to run locally on Apple Silicon and to be paired with a retrieval layer for current facts.

Unlike the smaller-base siblings, FR-Blaze sits on a strong 9B base that already scores ~85% on a public marketing benchmark. The adapter is a domain lens for voice and framing, not a capability boost. The capability is in the base; currency comes from retrieval. This card reports that honestly.

  Where a Thin Language Model fits

  specialization
     ▲
     │   ● FR-Blaze 9B
     │     marketing lens · local · grounded by retrieval
     │
     │                              ● Frontier LLM
     │                                broad · hosted · costly
     └──────────────────────────────────────────────▶ generality

How it works

  Marketing question
            │
            ▼
  ┌─────────────────────────────────────────────────────┐
  │ FR-Blaze 9B   (base capability + FR-Blaze lens)      │
  └─────────────────────────────────────────────────────┘
            │  + retrieved, dated facts (recommended)
            │
   advises across the digital stack
            │
            ├─ Search           SEO (technical/on-page/content/links) · Google & Microsoft Ads
            ├─ Paid social       Meta · TikTok · LinkedIn · YouTube · programmatic
            ├─ Organic social    strategy · content · community · cadence
            ├─ Content & email   briefs · distribution · lifecycle flows · deliverability
            └─ Analytics & CRO   GA4 · attribution · landing pages · budgeting · strategy
            │
            ▼
   Grounded, local answer  (verify live platform specs against current docs)

It is an assistant, not a certified authority, and not a substitute for verifying live platform features, ad specs, prices, or policies.

Specifications

Base	Gemma 2 9B (4-bit, MLX)
Parameters	~9.2B · 4-bit
Method	LoRA adapters, fused
Runtime	MLX (Apple Silicon)
Languages	English
License	Gemma Terms of Use

Quickstart

Runs locally with MLX on Apple Silicon. Apply a repetition penalty to avoid looping.

pip install -U mlx-lm
python3 -m mlx_lm generate \
  --model FahrenheitResearch/FR-Blaze-9B \
  --repetition-penalty 1.3 \
  --max-tokens 600 \
  --prompt "Plan a Q3 paid social test across Meta and TikTok for a DTC skincare brand."

Recommended generation settings: repetition_penalty = 1.3, temperature ≈ 0.5. A system prompt of "You are FR-Blaze, a marketing expert." matches training.

Recommended deployment: pair with retrieval. FR-Blaze is strongest when current, dated facts (platform updates, ad specs, account context) are retrieved and prepended to the prompt. The base model carries the stable knowledge; retrieval carries what changes. See the project for a reference retrieval setup.

Intended use

Assisting marketing, growth, and sales teams with everyday questions across the digital stack:

  FR-Blaze covers
  ├─ Search             SEO (technical, on-page, content, keywords, links), Google/Microsoft Ads, Shopping, PMax
  ├─ Paid social        Meta, TikTok, LinkedIn, YouTube, programmatic: targeting, creative, bidding, measurement
  ├─ Organic social     strategy, content, community, cadence, channel fit
  ├─ Content & email    briefs, formats, distribution, lifecycle flows, segmentation, deliverability
  └─ Analytics & CRO    GA4, attribution, incrementality, landing pages, budgeting, channel mix, strategy

Out of scope. Legal, financial, or compliance advice; autonomous spending or publishing; anything that must be exact (current platform specs, prices, policies) without grounding it in retrieved, current sources.

Evaluation

Scored on the public, MIT-licensed AdsGPT marketing benchmark (hand-authored against 2026 platform docs), deduped of its persona-templated copies to 377 unique multiple-choice questions across five marketing categories. Scoring is exact letter-match (no judge model), so it is fully reproducible. Decoding is deterministic (greedy).

  AdsGPT marketing benchmark, MCQ accuracy

  Critical thinking     ██████████████████░  91.7
  Email & lifecycle     █████████████████░░  86.2
  SEO & organic         ████████████████░░░  84.4
  Google Ads            ███████████████░░░░  78.8
  Meta Ads              ███████████████░░░░  78.8
  ──────────────────    ───────────────────  ────
  FR-Blaze-9B overall   ████████████████░░░  83.6

Category	FR-Blaze-9B	Gemma 2 9B base
Critical thinking	91.7%	91.7%
Email & lifecycle	86.2%	86.2%
SEO & organic	84.4%	84.4%
Google Ads	78.8%	83.8%
Meta Ads	78.8%	80.0%
Overall	83.6%	84.9%

Honest reading. The base model is already a strong marketing generalist (84.9%). The LoRA lens matches it on the stable categories and slightly trails on the two fast-changing paid-platform categories. The adapter adds FR-Blaze voice and framing, not benchmark capability. This is expected for a 9B base that already knows public marketing knowledge, and it is exactly why the recommended deployment grounds the model with retrieval for the volatile, current-fact categories rather than relying on the weights. Reproduce with the project's mcq_eval_9b.sh (it scores both the base and this model).

Limitations

Not a certified authority. Outputs assist research and drafting only. They are not a substitute for current platform documentation, financial sign-off, or legal advice.

• English only (v1).
• Lens, not a capability boost. On a public benchmark this model matches its base rather than beating it; its value is domain voice plus retrieval, not extra knowledge in the weights.
• Currency. Answers about live platform features, ad specs, prices, and policies are only as current as the facts you retrieve and provide. Without grounding, verify on the live platform.
• Self-reported, reproducible metric. The score is exact MCQ letter-match on a public benchmark; the harness is included so anyone can reproduce both base and tuned numbers.

Training

  Marketing instruction pairs ─┐
  SEO · paid search            │
  paid & organic social        ├─▶ LoRA fine-tune (MLX) ─▶ fuse adapters ─▶ FR-Blaze 9B
  email · analytics · math     │    gentle: lr 5e-5 · 4 layers · ~250 iters
  worked calculations          ┘

Base, method, data, and hyperparameters

• Base model: mlx-community/gemma-2-9b-it-4bit (Gemma 2 architecture, 4-bit)
• Method: LoRA adapters via mlx-lm (QLoRA on the 4-bit base), fused into this standalone model
• Data: ~200 curated marketing instruction pairs spanning SEO, paid search, paid and organic social, content, email, analytics/CRO, plus worked-math examples. Model-assisted synthetic pairs with human review; benchmark questions held out (no leakage, validator-enforced).
• Hyperparameters (gentle, to preserve base capability): learning rate 5e-5, LoRA layers 4, batch size 1, ~250 iters, max sequence length 1536, prompt masking on, gradient checkpointing on.
• Note: harder fine-tuning (higher lr, more layers/iters) degraded reasoning and math without adding capability; the gentle recipe above was chosen on benchmark evidence.

Citation

Gemma Terms of Use. Base model google/gemma-2-9b-it is governed by the Gemma license; this derivative carries the same terms.

@software{fr_blaze_2026,
  title  = {FR-Blaze-9B: a thin language model for online marketing},
  author = {Fahrenheit Research},
  year   = {2026},
  note   = {Fine-tuned from Gemma 2 9B (4-bit) with MLX/LoRA; deploy with retrieval}
}

---

FAHRENHEIT RESEARCH

Thin Language Models for specialized domains.

Website  ·  GitHub  ·  Sibling: FR-Forge-1.7B  ·  Sibling: FR-Lex-1.7B