Instructions to use AIHeikki/brainsurged-gladiator-gemma4-e2b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use AIHeikki/brainsurged-gladiator-gemma4-e2b with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M

Use Docker

docker model run hf.co/AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use AIHeikki/brainsurged-gladiator-gemma4-e2b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AIHeikki/brainsurged-gladiator-gemma4-e2b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AIHeikki/brainsurged-gladiator-gemma4-e2b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M

Ollama
How to use AIHeikki/brainsurged-gladiator-gemma4-e2b with Ollama:
```
ollama run hf.co/AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M
```
Unsloth Desktop

How to use AIHeikki/brainsurged-gladiator-gemma4-e2b with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use AIHeikki/brainsurged-gladiator-gemma4-e2b with Docker Model Runner:
```
docker model run hf.co/AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M
```

Lemonade

How to use AIHeikki/brainsurged-gladiator-gemma4-e2b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M

Run and chat with the model

lemonade run user.brainsurged-gladiator-gemma4-e2b-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use AIHeikki/brainsurged-gladiator-gemma4-e2b with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use AIHeikki/brainsurged-gladiator-gemma4-e2b with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "AIHeikki/brainsurged-gladiator-gemma4-e2b:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Brainsurged One-Shot Gemma Gladiator E2B (`v6-gladiator`)

Can a smaller open-weights model (google/gemma-4-E2B-it, 2.3B effective active text parameters vs. Qwen3.5-4B's 4.66B) learn a consolidated policy, agentic, and tactical chess curriculum in a single training run?

Base Model: google/gemma-4-E2B-it / unsloth/gemma-4-E2B-it (Apache-2.0, 35 layers, 2.3B active compute text parameters; Vision & Audio encoders stripped at GGUF export)
Flagship Artifact (main): brainsurged-gladiator-gemma4-e2b-Q4_K_M.gguf (3.19 GiB / 3,427,880,704 bytes, SHA-256: d28b4087b1c7386e2c1531c209ceeb3f5e46d6976ebd06cb7356c6ff302878ed)
Compared Against: AIHeikki/brainsurged-chessmaster-qwen3.5-4b (4.66B, 6 sequential surgery loops) and Jev 1.13.0 (api.typesafe.ai)

Interpretation and limits

This is a trained hybrid model: 2,780 synthetic examples and gradient-based LoRA, followed by zero-gradient calibration. It is not a zero-training-data or zero-gradient-only result.

The recorded two-game score was 1.5–0.5 against Chessmaster, and 0.5–1.5 against Jev. Four games and a 24-position sprint do not establish Elo or general superiority. Gladiator was faster than Jev in the two complete games, but slower in the sprint (214.35 vs 177.42 ms WAN p50). These are observations from those workloads, not universal latency claims.

The 2.3B figure is Gemma's effective parameter count; its embedding-inclusive total is larger (the base text model is approximately 5.1B). Comparing 2.3B effective parameters with Qwen's 4.66B total does not demonstrate half the parameter count or half the compute. No end-to-end FLOP measurement was performed.

“Zero leakage” below refers to the reported exact piece-placement overlap check only; it does not prove semantic independence or exclude benchmark-guided development. Holdout accuracy is specific to the supplied synthetic suites. The serving decoder, quantization, prompts and runtime are part of the evaluated system; stock Ollama/llama.cpp may produce different moves or latency. Parallel batch checks measure agreement on smoke-test cases with a numerical tolerance; they do not prove bit-exact agreement on every position.

Recorded Results: One Training Run and Workload-Specific Latency

Instead of accumulating 5–6 sequential rounds of LoRA + weight edits, Experiment 12 (v6-gladiator) consolidated our entire curriculum (2,780 zero-leakage rows: Round-4 policy/agentic + Round-6 Chessmaster + 380 new Queen/Rook ray & post-promotion contact-check guards) into one single surgery pass directly on raw Gemma 4 E2B-it, paired with serving_v7_gladiator (2-Pass Parallel Multi-seq_id Tree-Batch C++/CUDA engine).

1. Live Championship Chess Arena on Cloud Run `NVIDIA L4` (`python-chess` Referee + Stockfish Audit)

Full PGNs and ply-by-ply Stockfish evaluations archived in reports/gladiator_evaluations/live_matches/ and GLADIATOR_ARENA_REPORT.md.

Match Game	White	Black	Result	Plies	Termination	`≥300cp` Blunders (`W / B`)	Kamikaze Checks (`W / B`)	Engine `p50` (`WAN p50`)
Game 1 (vs. `Jev`)	`Jev 1.13.0`	`Gemma Gladiator E2B`	`½–½`	`113`	Insufficient material	`7 / 5`	`4 / 3`	`109.0 ms` (`161.2 ms`) vs. `Jev` `166.3 ms` WAN
Game 2 (vs. `Jev`)	`Gemma Gladiator E2B`	`Jev 1.13.0`	`0–1`	`72`	Checkmate (`36...Bh6#`)	`7 / 8`	`0` / `7`	`103.5 ms` (`155.0 ms`) vs. `Jev` `172.2 ms` WAN
Game 3 (vs. `4B`)	`Chessmaster 4B`	`Gemma Gladiator E2B`	`0–1`	`30`	Checkmate (`15...Qxc1#`)	`3 / 0`	`0 / 0`	`109.9 ms` (`179.1 ms`) vs. `4B` `205.5 ms` (`288.3 ms`)
Game 4 (vs. `4B`)	`Gemma Gladiator E2B`	`Chessmaster 4B`	`½–½`	`172`	Insufficient material	`21 / 18`	`9 / 4`	`127.5 ms` (`180.0 ms`) vs. `4B` `116.4 ms` (`171.4 ms`)

Defeated Brainsurged Chessmaster Qwen3.5-4B 1½–½ head-to-head:
- In Game 3 (Black), Gemma Gladiator E2B played a 0-blunder, 0-kamikaze-check 30-ply checkmate (15...Qxc1#) while running 1.87× faster on-GPU (109.9 ms vs. 205.5 ms engine p50).
- In Game 4 (White), Gemma Gladiator E2B promoted three separate passed pawns (39.e8=Q+, 56.h8=Q, 78.f8=Q) and held the draw (½–½), clinching the 1½–½ series win over the 4.66B 6-loop champion.
Sub-Jev Live WAN Latency (155.0–161.2 ms WAN p50):
- Across both live matches against Jev 1.13.0, Gemma Gladiator E2B beat Jev 1.13.0 on end-to-end WAN latency (155.0 ms vs. 172.2 ms in Game 2; 161.2 ms vs. 166.3 ms in Game 1), and committed fewer Stockfish ≥300cp blunders than Jev in both Game 1 (5 vs. 7) and Game 2 (7 with 0 kamikaze checks vs. Jev's 8 with 7 kamikaze checks, after an unsound 3.Nxf7 opening gambit).

2. Live `NVIDIA L4` 3-Way Stratified 24-Position Sprint (`Q4_K_M` GGUF + `serving_v7`)

Contender	Sprint Accuracy (`24`)	Trap Blunders	Engine `p50`	Engine Mean	WAN `p50`	Mean GPU Forward Passes
`One-Shot Gemma Gladiator E2B · v7 Parallel-Tree`	`75.00%`	`0.00%`	`160.19 ms`	`157.25 ms`	`214.35 ms`	`2.42` (`100%` parallel-tree)
`Brainsurged Chessmaster 4B · v7`	`70.83%`	`4.17%`	`260.84 ms`	`261.08 ms`	`339.73 ms`	`2.42`
`Jev 1.13.0` (`api.typesafe.ai`)	`83.33%`	`8.33%`	—	—	`177.42 ms`	—

3. `A100` Holdout Benchmark Suite (`Holdout-Chessmaster-120` & `Holdout-240`)

Source record: exp12_gladiator_campaign_results.json. Zero piece-placement FEN overlap (0/1,460) against both Holdout-Chessmaster-120 and all 331 Jevlab chess_lab FENs.

Model	`Holdout-Chessmaster-120` (`Zero-Think`)	`Holdout-Chessmaster-120` (`Micro-Blink`)	Trap Blunder Rate	Passed-Pawn (`30`)	Mate-in-1 (`30`)	Poisoned Check (`30`)	Queen Guard (`30`)	`Sequential-Agentic-120`	Combined `Holdout-240`
Raw `google/gemma-4-E2B-it`	`7.50%` (`0.83%` greedy)	`0.00%` (`100%` loop)	`30.83%`	`0.00%`	`0.00%`	`20.00%`	`10.00%`	—	—
`One-Shot Gemma Gladiator E2B` (`1` Loop)	`77.50%` (`+70.00 pp`)	`77.50%` (`29.1` tok)	`0.83%` (`0.0%` Blink)	`100.00%`	`100.00%`	`60.00%`	`50.00%`	`100.00%` (`0.0000` Brier)	`94.16%` (`0.0%` trap)
`Chessmaster Qwen3.5-4B` (`6` Loops)	`80.83%`	`80.83%`	`0.00%`	`100.00%`	`100.00%`	`66.67%`	`56.67%`	`100.00%`	`98.75%`

Anatomical Gemma 4 Surgery Design (`35` Layers, `18 KB/tok` KV Cache)

Why does Gemma 4 E2B hit sub-110 ms engine latency while absorbing a 6-round curriculum in a single loop?

Zone 1 — Frozen Sensory Stem & Per-Layer Embeddings (Layers 0..2 + embed_tokens_per_layer, 0.0000 Drift):
- Gemma 4 injects a fresh 256-dim Per-Layer Embedding (PLE) bypass into every layer (0..34). Freezing Layers 0..2 and the PLE lookup table bit-for-bit preserves raw JSON syntax and board-coordinate token grounding.
Zone 2 & 3 — Cortical Rank-24 LoRA (Layers 3..32, including Layers 20..34 Zero-New-KV Shared Fold):
- Gemma 4 alternates 4:1 512-token Sliding-Window Attention (head_dim=256, k_eq_v) with Global Attention (head_dim=512), and layers 20..34 (15 of 35 layers!) allocate zero new KV cache, reading directly from layers 18–19 (18 KB/token total KV footprint).
- Rank-24 LoRA (alpha=48, 234 linear modules, 33,951,744 scalars = 0.7282% of the language backbone) trained across 2,780 examples in 16.4 minutes on NVIDIA A100-40GB (loss: 2.5697 → 0.0547).
Zone 4 — Zero-Gradient Prefrontal Roof Calibration (Layers 33..34 + lm.norm, autograd=False):
- Accounts for Gemma 4's post_feedforward_layernorm, layer_scalar, and final_logit_softcapping=30.0, selecting beta_temp = 0.90 (0/16 unclosed <|channel>thought ... <channel|> loops, 30.9s wall clock).
Stage 3 — Multimodal Tower Stripping + serving_v7_gladiator 2-Pass Parallel-Tree Engine:
- Strips Vision & Audio towers during Q4_K_M GGUF conversion (3.19 GiB) and batches candidate move suffixes across parallel sequence branches. Pass counts depend on the position (2.42 mean in the recorded sprint).

Quickstart (Ollama / Stock `llama.cpp` / `serving_v7_gladiator`)

1. Run Locally with Ollama

ollama run hf.co/AIHeikki/brainsurged-gladiator-gemma4-e2b

2. Run with Stock `llama.cpp` (`llama-server`)

llama-server \
  --model brainsurged-gladiator-gemma4-e2b-Q4_K_M.gguf \
  --gpu-layers 99 \
  --flash-attn on \
  --jinja \
  --reasoning-budget 0

3. Build & Run `serving_v7_gladiator` 2-Pass Parallel-Tree CUDA Engine

See serving_v7_gladiator/README.md:

cd serving_v7_gladiator
python3 stage_models_from_hf.py
docker build --platform linux/amd64 -t jevlab-serving-v7-gladiator .

Repository Layout & SHA-256 Verification

Path	Artifact	Size	SHA-256
`brainsurged-gladiator-gemma4-e2b-Q4_K_M.gguf`	Flagship text-only `Q4_K_M` GGUF	`3.19 GiB` (`3,427,880,704` B)	`d28b4087b1c7386e2c1531c209ceeb3f5e46d6976ebd06cb7356c6ff302878ed`
`patches_and_attestations/v6/`	One-shot BF16 surgical delta `.safetensors` & JSON attestation	`136.8 MiB` (`143,453,552` B)	`brainsurged-gladiator-gemma4-e2b-oneshot-surgical-delta.safetensors`
`serving_v7_gladiator/`	`jev_jump_engine.cpp`, `fast_gateway.py`, `Dockerfile`, `smoke_test_build.py`	`~75 KiB`	2-Pass Parallel Multi-`seq_id` Tree-Batch CUDA serving stack
`reports/gladiator_evaluations/`	Live PGNs, `GLADIATOR_ARENA_REPORT.md`, `exp12_gladiator_campaign_results.json`	`~210 KiB`	Complete `A100` holdout, Stockfish blunder audit, and `L4` match records
`assets/gemma4_e2b_cortical_atlas.jpg`	35-layer Gemma 4 E2B anatomical cutaway diagram	`790 KiB`	Visual architecture reference

License & Attribution

Released under Apache-2.0, inheriting the license of google/gemma-4-E2B-it (Gemma 4 License).

Downloads last month: 4

GGUF

Model size

5B params

Architecture

gemma4

Hardware compatibility

4-bit

Model tree for AIHeikki/brainsurged-gladiator-gemma4-e2b

Base model

google/gemma-4-E2B

Finetuned

google/gemma-4-E2B-it

Adapter

(224)

this model

Brainsurged One-Shot Gemma Gladiator E2B (v6-gladiator)

Interpretation and limits

Recorded Results: One Training Run and Workload-Specific Latency

1. Live Championship Chess Arena on Cloud Run NVIDIA L4 (python-chess Referee + Stockfish Audit)

2. Live NVIDIA L4 3-Way Stratified 24-Position Sprint (Q4_K_M GGUF + serving_v7)

3. A100 Holdout Benchmark Suite (Holdout-Chessmaster-120 & Holdout-240)

Anatomical Gemma 4 Surgery Design (35 Layers, 18 KB/tok KV Cache)

Quickstart (Ollama / Stock llama.cpp / serving_v7_gladiator)

1. Run Locally with Ollama

2. Run with Stock llama.cpp (llama-server)

3. Build & Run serving_v7_gladiator 2-Pass Parallel-Tree CUDA Engine

Repository Layout & SHA-256 Verification

License & Attribution

Model tree for AIHeikki/brainsurged-gladiator-gemma4-e2b

Brainsurged One-Shot Gemma Gladiator E2B (`v6-gladiator`)

1. Live Championship Chess Arena on Cloud Run `NVIDIA L4` (`python-chess` Referee + Stockfish Audit)

2. Live `NVIDIA L4` 3-Way Stratified 24-Position Sprint (`Q4_K_M` GGUF + `serving_v7`)

3. `A100` Holdout Benchmark Suite (`Holdout-Chessmaster-120` & `Holdout-240`)

Anatomical Gemma 4 Surgery Design (`35` Layers, `18 KB/tok` KV Cache)

Quickstart (Ollama / Stock `llama.cpp` / `serving_v7_gladiator`)

2. Run with Stock `llama.cpp` (`llama-server`)

3. Build & Run `serving_v7_gladiator` 2-Pass Parallel-Tree CUDA Engine