Instructions to use jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge 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 jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge
# Run inference directly in the terminal:
llama cli -hf jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge
# Run inference directly in the terminal:
llama cli -hf jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge
# Run inference directly in the terminal:
./llama-cli -hf jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge
# Run inference directly in the terminal:
./build/bin/llama-cli -hf jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

Use Docker

docker model run hf.co/jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

LM Studio
Jan

vLLM

How to use jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

Ollama
How to use jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge with Ollama:
```
ollama run hf.co/jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge
```

Unsloth Studio

How to use jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge to start chatting

How to use jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

OpenClaw new

How to use jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

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 "jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge" \
  --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"

Docker Model Runner
How to use jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge with Docker Model Runner:
```
docker model run hf.co/jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge
```

Lemonade

How to use jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

Run and chat with the model

lemonade run user.Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

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 jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

Run Hermes

hermes

Atomic Chat

Qwen3.8-27B-CIRU-ActiveFPX-PromptForge v2

Qwen3.8-27B-CIRU-ActiveFPX-PromptForge v2 is Ciru's quality-and-speed release of Qwen3.8-27B for AMD ROCm systems. V2 upgrades the output projection to Q8 for better logit fidelity and coding quality while preserving essentially all of v1's throughput. It uses CIRU DualView: an ActiveFPX compact generation model plus PromptForge-optimized prompt views. Generation remains on the model's native target path and uses native multi-token prediction (MTP).

This release supports text and vision serving. The original Qwen3.8 vision path is preserved through a separate BF16 vision-projector companion.

Highlights

ActiveFPX — runtime-aware compact tensor allocation designed to retain model quality while reducing the memory and bandwidth cost of serving.
CIRU DualView — keeps the compact GGUF for generation and fallback while optional companion views accelerate large prompt operations.
PromptForge — the CIRU DualView execution engine, using prepacked projection views, merged projection work, fused FFN execution, fused SwiGLU-to-down packing, and merged Gated DeltaNet QKV/Z projection.
Native MTP — depth-4 native multi-token prediction for higher served generation throughput without replacing the model's own generation path.
Vision path preserved — the official vision tower and 5120-wide merger were converted into a separate BF16 projector companion without being modified by ActiveFPX or PromptForge.
Quality-aware v2 output — the final 248,320-token output projection uses Q8 to improve distribution fidelity and coding accuracy; embeddings and the native MTP module remain compatible with the Qwen3.8 architecture.
Serving defaults — 64K context, prompt cache enabled, continuous batching, temperature 0.7, top-p 0.95, and thinking disabled.
Fail-closed routing — PromptForge routes only the shapes and layers for which the accelerated path was built. Other work falls back to the native model path.

Benchmark highlights

All figures below were measured on one AMD Ryzen AI Max+ 395 / Radeon 8060S system using the pinned CIRU runtime. PP is prompt-processing throughput; TG is served generation throughput. Results are workload-specific and should not be treated as hardware-independent claims.

Highlight	Result
HumanEval pass@1	160 / 164 (97.56%)
HumanEval+ pass@1	151 / 164 (92.07%)
2,048-token PromptForge prefill	395.28 PP tok/s
8,192-token PromptForge prefill	356.98 PP tok/s
3,524-token prompt + 128-token native-MTP generation	353.54 PP tok/s / 38.33 served TG tok/s

Quality

Evaluation	Result	Notes
HumanEval pass@1	160 / 164 (97.56%)	Deterministic, chat-adapted EvalPlus execution; thinking off
HumanEval+ pass@1	151 / 164 (92.07%)	Same complete 164-problem run
HermesAgent-20	98 / 100	20 varied agent/tool tasks; cache on; temperature 0.7; top-p 0.95; thinking off; no retries
Compact-target KLD vs BF16	0.036162	Fixed paired calibration corpus
Compact-target perplexity change vs BF16	+2.79%	Fixed paired calibration corpus

What changed from v1

V2 changes the compact target's output projection to Q8. The PromptForge companion views, native MTP depth, context, cache policy, and public runtime configuration remain unchanged.

Matched artifact comparison	v1	v2	Change
HumanEval pass@1	157 / 164	160 / 164	+3 tasks
HumanEval+ pass@1	147 / 164	151 / 164	+4 tasks
KLD vs BF16	0.039182	0.036162	-7.71%
Same-top token agreement	89.951%	90.637%	+0.686 pp
Aggregate prompt processing	279.38 tok/s	276.72 tok/s	-0.95%
Aggregate served generation	40.59 tok/s	40.37 tok/s	-0.55%
HumanEval generation wall time	22m 09s	20m 34s	-7.18%

The v1-to-v2 table uses one complete, deterministic 164-problem comparison with identical runtime and serving settings for both artifacts. V2 generated 8.34% fewer tokens, which is why end-to-end wall time improved despite nearly identical measured token throughput.

The HumanEval figures use a chat-adapted serving protocol and are not presented as canonical raw-completion leaderboard submissions.

PromptForge performance

Prompt workload	CIRU release	Matched compact path	Change
2,048-token prompt	395.28 PP tok/s	325.22 PP tok/s	+21.54%
8,192-token prompt	356.98 PP tok/s	296.23 PP tok/s	+20.51%
3,524-token prompt + 128 generated	353.54 PP tok/s, 38.33 TG tok/s	334.66 PP tok/s, 40.95 TG tok/s	+5.64% PP, -6.40% TG, -2.55% total wall time

The isolated prompt rows show the PromptForge route's strongest measured prefill gains. The end-to-end row includes native MTP generation and is the more representative latency comparison.

HermesAgent-20

Result	Measurement
Agent/tool score	98 / 100
Prompt processing	310.24 PP tok/s
Served generation	33.68 TG tok/s

This run used 20 varied agent/tool tasks with prompt cache enabled, temperature 0.7, top-p 0.95, thinking disabled, and no retries.

CIRU DualView: speed or memory

The companion views are optional. Enable them for maximum prompt throughput, or run the GGUF alone to save 21.15 GB / 19.70 GiB of storage and shared memory.

Workload	DualView on	DualView off	PP gain
2,048-token prompt	395.28 tok/s	325.22 tok/s	+21.54%
8,192-token prompt	356.98 tok/s	296.23 tok/s	+20.51%
3,524-token end-to-end prompt	353.54 tok/s	334.66 tok/s	+5.64%
HermesAgent-20 prompts	310.24 tok/s	303.62 tok/s	+2.18%

DualView gains depend on prompt shape and how much of the request uses the accelerated physical prompt blocks. Cache hits and small or irregular tails reduce the percentage improvement.

Vision support

Vision is supported through the matching BF16 projector companion, which contains the Qwen3.8 vision tower and 5120-wide merger. ActiveFPX changes the language-model tensor allocation, while PromptForge accelerates selected language-model prompt operations; neither rewrites the vision tensors. Unsupported PromptForge shapes continue through the native runtime path.

Pass the companion to the CIRU runtime with --mmproj /path/to/vision-projector.mmproj.

Release files

Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-v2.gguf — recommended v2 model artifact with Q8 output projection
Qwen3.8-27B-CIRU-ActiveFPX-PromptForge.gguf — original v1 model artifact, retained for reproducibility
Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-FFN.pfs — PromptForge FFN companion
Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-GDN.pfs — PromptForge Gated DeltaNet companion
assets/qwen3827bactivefpx.png — model artwork

The PromptForge companion files are required only for CIRU DualView acceleration. The GGUF remains independently usable through the native runtime path.

Pinned runtime

This release requires the CIRU ROCmFPX runtime at:

Repository: ciru-ai/ROCmFPX
Release tag: qwen3.8-activefpx-promptforge-v2
Runtime commit: 78e97d1e5243245a5819ead97d74ccf1770dfbf6
AMD Composable Kernel commit: fdf4bb7fcc984811cef48ce817d89aac064b984a

Build and dependency instructions are maintained in docs/activefpx-promptforge-qwen38.md.

Serving examples

CIRU DualView on — maximum prompt speed

export PROMPTFORGE_SIDECAR=/models/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-FFN.pfs
export PROMPTFORGE_GDN_SIDECAR=/models/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-GDN.pfs
export PROMPTFORGE_MODE=m2048_fused_tail1476
export GGML_CUDA_GRAPH_OPT=0

./llama-server \
  -m /models/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-v2.gguf \
  --alias main --host 127.0.0.1 --port 8080 \
  -dev ROCm0 -ngl 999 -c 65536 -b 2048 -ub 2048 \
  -fa on -ctk f16 -ctv f16 -np 1 -ctxcp 0 \
  --cache-ram 8192 --cache-prompt --metrics \
  --spec-type draft-mtp --spec-draft-device ROCm0 --spec-draft-ngl 999 \
  --spec-draft-type-k f16 --spec-draft-type-v f16 \
  --spec-draft-n-max 4 --spec-draft-n-min 0 \
  --spec-draft-p-min 0.0 --spec-draft-p-split 0.10 \
  --spec-draft-backend-sampling \
  --temp 0.7 --top-p 0.95 \
  --reasoning off --reasoning-format none --reasoning-budget -1

CIRU DualView off — save 19.70 GiB

Do not download the two .pfs files, or remove their environment variables before launching:

unset PROMPTFORGE_SIDECAR
unset PROMPTFORGE_GDN_SIDECAR
unset PROMPTFORGE_MODE

./llama-server \
  -m /models/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-v2.gguf \
  --alias main --host 127.0.0.1 --port 8080 \
  -dev ROCm0 -ngl 999 -c 65536 -b 2048 -ub 2048 \
  -fa on -ctk f16 -ctv f16 -np 1 -ctxcp 0 \
  --cache-ram 8192 --cache-prompt --metrics \
  --spec-type draft-mtp --spec-draft-device ROCm0 --spec-draft-ngl 999 \
  --spec-draft-type-k f16 --spec-draft-type-v f16 \
  --spec-draft-n-max 4 --spec-draft-n-min 0 \
  --spec-draft-p-min 0.0 --spec-draft-p-split 0.10 \
  --spec-draft-backend-sampling \
  --temp 0.7 --top-p 0.95 \
  --reasoning off --reasoning-format none --reasoning-budget -1

This keeps ActiveFPX and native MTP but disables PromptForge's additional prompt views. V2 size falls from 37.46 GB / 34.89 GiB for the complete three-file payload to 16.31 GB / 15.19 GiB for the GGUF alone. KV cache and runtime overhead are additional in both modes.

Memory requirements vary with context, cache size, and platform. Start with one slot and adjust only after measuring available host and GPU-visible memory.

Architecture and intended use

The base model is Qwen3.8-27B: a 27B language model in a 28B multimodal architecture with hybrid attention/Gated DeltaNet layers and a native MTP module. This release supports text and vision serving for local inference, coding, agent workflows, general assistants, and research.

Users remain responsible for application-specific evaluation, safety controls, and compliance. Benchmark scores do not guarantee behavior on unseen tasks or languages.

Credits

License

The derived model artifact follows the Qwen3.8-27B Apache-2.0 license. Runtime source and third-party components retain their respective licenses; consult the pinned runtime repository for details.

Downloads last month: -

GGUF

Model size

27B params

Architecture

qwen35

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge

Base model

Qwen/Qwen3.8-27B

Quantized

(401)

this model