README.md

---
license: apache-2.0
base_model:
- stepfun-ai/Step-3.7-Flash
- notSnix/Step-3.7-Flash-MTP-Draft-GGUF
base_model_relation: quantized
language:
- en
library_name: llama.cpp
pipeline_tag: text-generation
tags:
- stepfun
- step3p7
- step-3.7
- step-3.7-flash
- gguf
- mtp
- speculative-decoding
- rocm
- vulkan
- rocmfpx
- fpx3
- q3
- q3_0_rocmfpx
- qualityplus
- amd
- ryzen-ai-max-395
- strix-halo
- agentic
- tool-calling
- long-context
---

![StepFun Step 3.7 ROCmFPX Q3 QualityPlus](assets/stepfun.png)

# Step-3.7-Flash-ROCmFPX-Q3-QualityPlus

This is an extremely high quality FPX3 / ROCmFPX Q3 GGUF build of `stepfun-ai/Step-3.7-Flash`, tuned for AMD Strix Halo local serving with Step MTP.

The goal is simple: keep Step 3.7 Flash useful at 256K context, keep the quality as high as possible, and keep it as small as possible. This release is a true tight Q3-weight build: `3.57 BPW`, `81.77 GiB` of language-model shards, and strong agent/tool behavior in local evals.

Use this if you want the Step 3.7 behavior profile, MTP support, and a much smaller local footprint than the stock GGUF Q3_K_L or ROCmFP4 STRIX_LEAN builds.

> **Required runtime:** these GGUFs do **not** run on stock upstream llama.cpp. They use ROCmFPX tensor types such as `q3_0_rocmfpx` plus Chadrock/ROCmFPX serving support for Step MTP. Build the pinned Ciru ROCmFPX runner below before trying to load the model.

## Why This One

Step 3.7 is huge. The practical local problem is not only speed; it is fitting enough context, KV, and agent workload into memory.

This FPX3/Q3 QualityPlus recipe was built for that constraint:

- `3.57 BPW` effective language-model size
- `81.77 GiB` total language GGUF shards
- `16.31%` smaller than the local ROCmFP4 STRIX_LEAN build
- `14.35%` smaller than StepFun's original `Q3_K_L` GGUF split
- up to 256K one-slot serving profile with q8_0 target KV and q8_0 draft KV
- Step MTP Q8 draft support through `draft-mtp`
- downloadable fixed Step tool/chat template using native `tool_response` observations and protocol-boundary escaping

In practice, the original StepFun `Q3_K_L` local split was not a compact 3-bit-feeling model: it measured about `95.46 GiB`, or roughly `4.17 BPW` by effective size. This QualityPlus build is the one I would publish/use as the FPX3 lane.

## Size Comparison

Measured from local GGUF shards:

| Build | Effective BPW | Shard total | Difference vs this release |
| --- | ---: | ---: | ---: |
| **ROCmFPX Q3 QualityPlus** | `3.57 BPW` | `81.77 GiB` | baseline |
| StepFun original `Q3_K_L` | `~4.17 BPW` | `95.46 GiB` | `+13.70 GiB` larger |
| ROCmFP4 STRIX_LEAN | `~4.27 BPW` | `97.70 GiB` | `+15.93 GiB` larger |

That size gap matters because Step 3.7 needs memory for long context, q8 KV, and MTP draft state. On the tested Strix Halo host, the Q3 QualityPlus 64K MTP profile used about `96.3 GiB` peak pooled GPU memory during long tool/Hermes runs, leaving enough RAM headroom to run the evals cleanly.

## Quality Highlights

This is not a throwaway low-bit build. The recipe protects the tensors that were most important for behavior while pushing the giant expert FFN tensors into `q3_0_rocmfpx`.

Local quality results on AMD Ryzen AI Max+ 395 / Strix Halo:

| Benchmark | Result | Notes |
| --- | ---: | --- |
| Tool-Eval full, 69 scenarios | [`88/100`, `122/138` raw points](evals/tool-eval-q3-qualityplus.json) | Same headline score as the recorded Step ROCmFP4 tool-eval row |
| HermesAgent-20, best Q3 run | `85/100` | `13.40 min`, `35.31 tok/s` decode, `96.37 GiB` peak pooled GPU |

The best recorded Q3 HermesAgent-20 run was very close to the local BF16 Qwen3.6 27B MTP reference row:

| Model / row | HermesAgent-20 score | Wall time |
| --- | ---: | ---: |
| BF16 Qwen3.6 27B MTP GGUF | `87/100` | `42.4 min` |
| **Step 3.7 ROCmFPX Q3 QualityPlus** | `85/100` | `13.4 min` |

That is within two points of the BF16 Qwen3.6 27B row on the local HermesAgent-20 suite, while running in a much more compact Step 3.7 Q3 package.

Exact Q3 QualityPlus tool-eval score summary: [`evals/tool-eval-q3-qualityplus.json`](evals/tool-eval-q3-qualityplus.json). Public reference page for the Step 3.7 tool-calling work: [StepFun Step 3.7 Tool Eval on llm.ciru.ai](https://llm.ciru.ai/tooleval/stepfun-step37-rocmfp4-mtp-vulkan-64k/). The Q3 QualityPlus full run used the same 69-scenario tool-eval harness and scored `88/100` locally.

## Speed

Q3 QualityPlus speed was effectively tied with the local ROCmFP4 Step build while using much less disk space.

Short-context MTP speed, Vulkan0, q8_0/q8_0 target KV, q8_0/q8_0 draft KV, one slot, `n_max=2`, `p_min=0.75`, `b8192/u2048`, 128 generated tokens:

| Prompt | PP tok/s | TG tok/s |
| ---: | ---: | ---: |
| `2k` | `309.44` | `29.97` |
| `4k` | `325.18` | `29.39` |
| `8k` | `311.15` | `28.58` |
| `16k` | `306.37` | `26.26` |

Compared with the local ROCmFP4 Step build:

| Prompt | Q3 QualityPlus TG | ROCmFP4 TG | Takeaway |
| ---: | ---: | ---: | --- |
| `2k` | `29.97` | `26.52` | Q3 faster |
| `4k` | `29.39` | `29.37` | tied |
| `8k` | `28.58` | `28.02` | tied/slightly Q3 |
| `16k` | `26.26` | `26.42` | tied |

128K stress row:

| Context | PP tok/s | TG tok/s | Peak pooled GPU |
| ---: | ---: | ---: | ---: |
| `~130k prompt` | `146.67` | `14.52` | `~95.36 GiB` |

At 128K, MTP initialized but produced no accepted drafts in that particular row, so treat the 128K decode number as an effective no-draft long-context decode reference.

256K load proof:

| Context | Proof | Memory state |
| ---: | --- | ---: |
| `262144` | target + Q8 MTP draft loaded, one slot, `draft-mtp`, `/v1/models` reports `n_ctx=262144` and `n_ctx_train=262144` | `~99.04 GiB` pooled GPU used, `~16 GiB` system RAM available |

The 256K row is a load/allocation proof, not a 256K prompt prefill benchmark.

## Files

Published shard names intentionally match the model name:

```text
Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00001-of-00009.gguf
Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00002-of-00009.gguf
Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00003-of-00009.gguf
Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00004-of-00009.gguf
Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00005-of-00009.gguf
Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00006-of-00009.gguf
Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00007-of-00009.gguf
Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00008-of-00009.gguf
Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00009-of-00009.gguf
```

The Step MTP draft model is not duplicated here. If you enable `draft-mtp`, you
must also download and pass the separate Q8 draft from
[`notSnix/Step-3.7-Flash-MTP-Draft-GGUF`](https://huggingface.co/notSnix/Step-3.7-Flash-MTP-Draft-GGUF),
for example `Step-3.7-Flash-MTP-Q8_0.gguf`. The main Q3 target GGUF does not
contain the MTP draft layers.

This repo also includes the tested chat/tool template:

```text
step37-native-tool-response-template.jinja
```

Download the target shards and template:

```bash
huggingface-cli download jcbtc/Step-3.7-Flash-ROCmFPX-Q3-QualityPlus \
  --include "Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-*.gguf" \
  --include "step37-native-tool-response-template.jinja" \
  --local-dir /mnt/models/jcbtc-Step-3.7-Flash-ROCmFPX-Q3-QualityPlus
```

Download the required Q8 MTP draft:

```bash
huggingface-cli download notSnix/Step-3.7-Flash-MTP-Draft-GGUF \
  Step-3.7-Flash-MTP-Q8_0.gguf \
  --local-dir /mnt/models/notSnix-Step-3.7-Flash-MTP-Draft-GGUF
```

Direct template URL:

```text
https://huggingface.co/jcbtc/Step-3.7-Flash-ROCmFPX-Q3-QualityPlus/resolve/main/step37-native-tool-response-template.jinja
```

Direct Q8 draft URL:

```text
https://huggingface.co/notSnix/Step-3.7-Flash-MTP-Draft-GGUF/resolve/main/Step-3.7-Flash-MTP-Q8_0.gguf
```

## Required ROCmFPX Runner

This model is tied to the Charlie/Ciru ROCmFPX llama.cpp runner family. A stock `llama-server` will not understand the ROCmFPX tensor types in these shards and will not reproduce the MTP serving behavior used for the benchmark rows.

Use the pinned Ciru runner:

```text
repo: https://github.com/ciru-ai/ROCmFPX
current recommended pin: 221402af8574faf652b101b6afe225a3f329561f
branch at time of pin: main
upstream lineage: charlie12345/ROCmFPX
```

The earlier Chadrock v2 speed-runner tag remains useful for historical comparison:

```text
tag: chadrockv2-runner-20260622
commit: 7aa484a2f0a504dc612a3d74a068024f3e6d6353
```

The Q3 QualityPlus Step 3.7 rows on this card were validated with the Chadrock/ROCmFPX runner path on AMD Ryzen AI Max+ 395 / Strix Halo. For fresh installs, use the current Ciru pin above unless you are reproducing an older benchmark exactly.

Build the runner on a Linux system with a working ROCm/HIP toolchain, Vulkan
development headers, CMake, and a C++ compiler. This is the pinned Strix Halo
reference build used by Ciru; it is not a universal distro installer, so package
names and ROCm paths may differ on Ubuntu, Arch, Fedora, NixOS, and other
distros.

```bash
git clone https://github.com/ciru-ai/ROCmFPX.git
cd ROCmFPX
git checkout 221402af8574faf652b101b6afe225a3f329561f

env JOBS="$(nproc)" \
  CMAKE_HIP_ARCHITECTURES=gfx1151 \
  ROCMFPX_DECODE_TUNE=stable \
  scripts/build-strix-rocmfp4-mtp.sh llama-server llama-bench
```

If your ROCm or rocWMMA headers live outside the script defaults, set the relevant environment variables before running the build, for example `ROCM_WMMA_INCLUDE=/path/to/rocWMMA/library/include`. If your GPU is not Strix Halo / gfx1151, change `CMAKE_HIP_ARCHITECTURES` for your target.

The script and build directory still use the historical `rocmfp4` name, but this is the ROCmFPX/Chadrock runner. For this model, the required support is ROCmFPX Q3 tensor support, not a ROCmFP4-only runtime.

The server binary should be:

```text
./build-strix-rocmfp4/bin/llama-server
```

Again, `build-strix-rocmfp4` is the historical build-directory name used by the ROCmFPX runner script.

If the model load fails with an unknown GGUF tensor type, you are using the wrong runner.

## Recommended Serving Profile

The locally tested long-context profile:

```text
context: up to 262144
slots: 1
backend: Vulkan0 target + Vulkan0 draft
MTP: --spec-type draft-mtp
draft model: Step-3.7-Flash-MTP-Q8_0.gguf from notSnix/Step-3.7-Flash-MTP-Draft-GGUF
speculative.n_max: 2
speculative.n_min: 0
speculative.p_min: 0.75
speculative.p_split: 0.10
batch / ubatch: 8192 / 2048
target KV: q8_0 / q8_0
draft KV: q8_0 / q8_0
prompt cache: disabled for 256K fit runs
sampler: temperature 1.0, top_p 0.95, min_p 0.0, repeat_penalty 1.0
reasoning: on, DeepSeek format
chat template: Step native tool_response template with protocol-boundary escaping
```

### Temperature Guidance

For coding agents, tool calls, patches, exact filenames, JSON, and reading source files, use a lower temperature than the general profile above:

```text
temperature: 0.0 to 0.1
top_p: 1.0
min_p: 0.0
repeat_penalty: 1.0
```

Use `temperature: 0.0` when exactness and reproducibility matter most, or `0.1` for a small amount of flexibility. Higher temperatures such as `0.7` to `1.0` are better reserved for brainstorming, varied prose, and creative exploration; they can make Step less reliable when reproducing identifiers or interpreting long tool/file output.

Serving backend note: on the tested AMD Ryzen AI Max+ 395 / Strix Halo system, this Step 3.7 Q3 build worked best through the ROCmFPX/Chadrock runner serving on `Vulkan0` for both target and draft. In the command below, ROCmFPX is the required tensor/runtime support; `-dev Vulkan0` and `--spec-draft-device Vulkan0` are the recommended serving backend.

For `models.ini`-style launchers, make sure the draft path is present. Setting
`spec-type = draft-mtp` without `spec-draft-model` makes the runner try to build
an MTP draft context from the main target GGUF, which fails because the target
does not contain MTP draft layers.

```ini
model = /mnt/models/jcbtc-Step-3.7-Flash-ROCmFPX-Q3-QualityPlus/Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00001-of-00009.gguf
chat-template-file = /mnt/models/jcbtc-Step-3.7-Flash-ROCmFPX-Q3-QualityPlus/step37-native-tool-response-template.jinja

spec-type = draft-mtp
spec-draft-model = /mnt/models/notSnix-Step-3.7-Flash-MTP-Draft-GGUF/Step-3.7-Flash-MTP-Q8_0.gguf
spec-draft-device = Vulkan0
spec-draft-ngl = all
spec-draft-type-k = q8_0
spec-draft-type-v = q8_0
spec-draft-n-max = 2
spec-draft-n-min = 0
spec-draft-p-min = 0.75
spec-draft-p-split = 0.10
```

If you see `context type MTP requested but model doesn't contain MTP layers`,
the draft model is missing or the path is wrong.

Example shape:

```bash
./build-strix-rocmfp4/bin/llama-server \
  -m Step-3.7-Flash-ROCmFPX-Q3-QualityPlus-00001-of-00009.gguf \
  --alias step-3.7-flash-rocmfpx-q3-qualityplus \
  --host 127.0.0.1 \
  --port 8080 \
  --jinja \
  -c 262144 \
  --reasoning on \
  --reasoning-format deepseek \
  --reasoning-budget -1 \
  -dev Vulkan0 \
  -ngl 999 \
  -fa on \
  -b 8192 \
  -ub 2048 \
  --parallel 1 \
  --no-mmap \
  --cache-ram 0 \
  -ctk q8_0 \
  -ctv q8_0 \
  --spec-draft-model Step-3.7-Flash-MTP-Q8_0.gguf \
  --spec-draft-device Vulkan0 \
  --spec-type draft-mtp \
  --spec-draft-ngl all \
  --spec-draft-type-k q8_0 \
  --spec-draft-type-v q8_0 \
  --spec-draft-n-max 2 \
  --spec-draft-n-min 0 \
  --spec-draft-p-min 0.75 \
  --spec-draft-p-split 0.10 \
  --chat-template-file /path/to/step37-native-tool-response-template.jinja \
  --metrics
```

## Template Note

The best local Step setup uses the included [`step37-native-tool-response-template.jinja`](step37-native-tool-response-template.jinja) template. It renders tool outputs as `tool_response` turns and escapes protocol-boundary tokens inside tool output. This is a general protocol-adapter fix: tool/file/search results stay observations instead of being flattened into user text.

Download:

```bash
curl -L -o step37-native-tool-response-template.jinja \
  https://huggingface.co/jcbtc/Step-3.7-Flash-ROCmFPX-Q3-QualityPlus/resolve/main/step37-native-tool-response-template.jinja
```

That matters for real agents because Step 3.7 can otherwise confuse tool output with conversation authority, especially in file/search-result injection cases.

## Build Notes

These are model-build notes, not runner-build instructions. Build the pinned ROCmFPX runner in the section above before serving the GGUFs.

The QualityPlus policy used here:

- huge `ffn_*_exps` tensors: `q3_0_rocmfpx`
- attention q/output protected at `q5_K`
- attention k/v protected at `q4_K`
- shared/dense FFN protected at `q5_K`
- output/token embeddings at `q4_0_rocmfp4_fast`

Converter-reported size: `83726.08 MiB / 3.57 BPW`, 9 shards.

## Credits

- Base model: [`stepfun-ai/Step-3.7-Flash`](https://huggingface.co/stepfun-ai/Step-3.7-Flash)
- MTP draft GGUF source: [`notSnix/Step-3.7-Flash-MTP-Draft-GGUF`](https://huggingface.co/notSnix/Step-3.7-Flash-MTP-Draft-GGUF)
- ROCmFPX creator: Charlie, `charlie12345` / `@italianclownz`, [`charlie12345/ROCmFPX`](https://github.com/charlie12345/ROCmFPX)
- Pinned public runner fork and build recipe: [`ciru-ai/ROCmFPX`](https://github.com/ciru-ai/ROCmFPX), current recommended pin `221402af8574faf652b101b6afe225a3f329561f`
- Quantization, the ROCmFPX Step 3.7 Q3 QualityPlus recipe, Strix Halo profile, and local benchmark work: Crown / Ciru

## Caveats

- This is a custom ROCmFPX GGUF release. It requires the compatible ROCmFPX/Chadrock llama.cpp runner; stock llama.cpp is not expected to load it.
- Quality numbers are local Strix Halo measurements and depend on runtime, chat template, KV type, and MTP settings.
- The model is strong but not perfect at autonomous email/message side effects; it can be cautious and ask for subject/body/recipient details instead of sending with inferred defaults.