stepfun-ai/Step-3.7-Flash-GGUF
Summary
StepFun releases GGUF quantizations of their 198B-parameter sparse MoE vision-language model Step-3.7-Flash, enabling local deployment with up to 256K context and selectable reasoning levels.
View Cached Full Text
Cached at: 06/01/26, 08:22 AM
stepfun-ai/Step-3.7-Flash-GGUF · Hugging Face
Source: https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF [ModelPage]:https://static.stepfun.com/blog/step-3.7-flash/
https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF#1-introduction1. Introduction
GGUF quantizations ofstepfun\-ai/Step\-3\.7\-Flash.
Step-3.7-Flash is a 198B-parameter sparse Mixture-of-Experts vision-language model from StepFun-ai, activating ~11B parameters per token for up to 400 t/s throughput. It pairs a 196B-parameter language backbone with a 1.8B-parameter vision encoder for native image understanding, supports a 256K context window, and offers three selectable reasoning levels (low / medium / high) to balance speed, cost, and depth. Built for agentic workloads — tool calling, multi-step reasoning, code, and math — with native multilingual coverage.
A separatemmprojprojector ships alongside the language quants for multimodal inference. With 128 GB of unified memory (Mac Studio, DGX Spark, Ryzen AI Max+ 395, etc.), you can privately host Step-3.7-Flash: Q4 quants and below run at full 256K context with high precision.
https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF#2-files2. Files
FileQuantSizeNotesStep\-3\.7\-flash\-BF16\.ggufBF16394 GBFull-precision reference.Step\-3\.7\-flash\-Q8\_0\.ggufQ8_0209 GBNear-lossless. Doesnotuse imatrix.Step\-3\.7\-flash\-Q4\_K\_S\.ggufQ4_K_S112 GBimatrix-calibrated. Balanced quality / size.Step\-3\.7\-flash\-IQ4\_XS\.ggufIQ4_XS105 GBimatrix-calibrated. Slightly smaller than Q4_K_S, comparable quality.Step\-3\.7\-flash\-Q3\_K\_L\.ggufQ3_K_L103 GBimatrix-calibrated. Aggressive size reduction.Step\-3\.7\-flash\-Q3\_K\_M\.ggufQ3_K_M94 GBimatrix-calibrated. Use when you need to fit on a single 64-96 GB device; expect modest quality loss at low bit-widths.Step\-3\.7\-flash\-IQ3\_XXS\.ggufIQ3_XXS76 GBimatrix-calibrated. Recommended only when memory is the primary constraint; offers the smallest footprint among the provided quantizations.mmproj\-Step\-3\.7\-flash\-f16\.ggufF164 GBVision projector. Pair with any of the language quants above for image input.
https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF#3-quickstart3. Quickstart
Build llama.cpp and run:
# 1. Clone and build
git clone https://github.com/stepfun-ai/llama.cpp.git
cd llama.cpp
git checkout -b step3.7 origin/step3.7
cmake -B build -DLLAMA_BUILD_TOOLS=ON -DLLAMA_BUILD_SERVER=ON
cmake --build build --config Release -j$(nproc)
# 2. Test performance (benchmark)
./build/bin/llama-batched-bench \
-m Step-3.7-flash-Q4_K_S.gguf \
-c 32768 -b 2048 -ub 2048 \
-npp 0,2048,8192,16384,32768 -ntg 128 -npl 1
# 3. Text-only inference
./build/bin/llama-cli \
-m Step-3.7-flash-Q4_K_S.gguf \
-c 32768 -ngl 99 -fa on \
-p "Write a Python function to compute the n-th Fibonacci number."
# 4. With vision (image + text)
./build/bin/llama-mtmd-cli \
-m Step-3.7-flash-Q4_K_S.gguf \
--mmproj mmproj-Step-3.7-flash-f16.gguf \
-c 32768 -ngl 99 -fa on \
--image path/to/image.jpg \
-p "Describe this image."
# 5. OpenAI-compatible server (text + vision)
./build/bin/llama-server \
-m Step-3.7-flash-Q4_K_S.gguf \
--mmproj mmproj-Step-3.7-flash-f16.gguf \
-c 32768 -ngl 99 -fa on \
--host 0.0.0.0 --port 8080
For full CLI / server options, see thellama.cpp README.
https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF#4-performance4. Performance
https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF#apple-mac-studio-m4-max-128-gb-unified-memoryApple Mac Studio (M4 max, 128 GB unified memory)
Step-3.7-flash-Q4_K_S
./llama-batched-bench -m Step-3.7-flash-Q4_K_S.gguf -c 262150 -b 2048 -ub 1024 -npp 0,2048,8192,16384,32768,65536,131072,262144 -ntg 128 -npl 1
PPTGPLN_KVT_PP sS_PP t/sT_TG sS_TG t/sT sS t/s012811280.0000.002.50051.202.50051.202048128121764.873420.282.63948.517.512289.6881921281832020.292403.702.75746.4323.049360.971638412811651242.854382.322.92443.7745.779360.693276812813289695.168344.323.22339.7298.391334.3465536128165664233.885280.213.90932.74237.794276.141310721281131200635.499206.255.75922.23641.258204.6026214412812622722362.488110.9613.1889.712375.677110.40 Step-3.7-flash-IQ4_XS
./llama-batched-bench -m Step-3.7-flash-IQ4_XS.gguf -c 262150 -b 2048 -ub 1024 -npp 0,2048,8192,16384,32768,65536,131072,262144 -ntg 128 -npl 1
PPTGPLN_KVT_PP sS_PP t/sT_TG sS_TG t/sT sS t/s012811280.0000.002.58249.582.58249.582048128121764.835423.562.67947.787.514289.6081921281832019.954410.552.80345.6622.757365.601638412811651242.142388.782.95743.2945.098366.133276812813289693.489350.503.28838.9396.777339.9165536128165664227.088288.593.94532.44231.033284.221310721281131200635.047206.405.79122.10640.838204.7326214412812622722170.271120.7913.0709.792183.342120.12 Step-3.7-flash-Q3_K_L
./llama-batched-bench -m Step-3.7-flash-Q3_K_L.gguf -c 262272 -b 2048 -ub 1024 -npp 0,2048,8192,16384,32768,65536,131072,262144 -ntg 128 -npl 1
PPTGBN_KVT_PP sS_PP t/sT_TG sS_TG t/sT sS t/s012811280.0000.003.59035.663.59035.662048128121765.263389.153.70234.578.965242.7281921281832021.789375.973.81733.5325.606324.921638412811651245.819357.583.97732.1849.796331.5932768128132896100.827324.994.30829.71105.135312.8965536128165664242.172270.624.97725.72247.149265.691310721281131200659.645198.706.76418.92666.409196.8826214412812622722200.370119.1414.0089.142214.378118.44
https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF#nvidia-dgx-spark-gb10-128-gb-unified-memoryNVIDIA DGX Spark (GB10, 128 GB unified memory)
Step-3.7-flash-Q4_K_S
./llama-batched-bench -m Step-3.7-flash-Q4_K_S.gguf -c 131300 -b 2048 -ub 1024 -npp 0,2048,8192,16384,32768,65536,131072 -ntg 128 -npl 1
PPTGBN_KVT_PP sS_PP t/sT_TG sS_TG t/sT sS t/s012811280.0000.005.15724.825.15724.822048128121768.021255.334.90726.0812.929168.3181921281832010.866753.895.16924.7616.035518.861638412811651229.389557.496.21520.6035.603463.783276812813289652.501624.146.93118.4759.432553.5065536128165664112.321583.477.76916.48120.090546.791310721281131200281.479465.669.83413.02291.313450.37 Step-3.7-flash-IQ4_XS
./llama-batched-bench -m Step-3.7-flash-IQ4_XS.gguf -c 262272 -b 2048 -ub 1024 -npp 0,2048,8192,16384,32768,65536,131072,262144 -ntg 128 -npl 1
PPTGPLN_KVT_PP sS_PP t/sT_TG sS_TG t/sT sS t/s012811280.0000.005.36823.855.36823.852048128121764.250481.875.31124.109.561227.5881921281832012.531653.735.81722.0118.348453.461638412811651224.474669.445.91521.6430.389543.353276812813289651.976630.446.53119.6058.508562.2565536128165664116.305563.487.93416.13124.239528.531310721281131200298.746438.7410.26312.47309.009424.582621441281262272924.872283.4414.8628.61939.734279.09 Step-3.7-flash-Q3_K_L
./llama-batched-bench -m Step-3.7-flash-Q3_K_L.gguf -c 262272 -b 2048 -ub 1024 -npp 0,2048,8192,16384,32768,65536,131072,262144 -ntg 128 -npl 1
PPTGPLN_KVT_PP sS_PP t/sT_TG sS_TG t/sT sS t/s012811280.0000.005.94721.525.94721.522048128121764.145494.085.62322.769.768222.7781921281832014.889550.205.79922.0720.688402.171638412811651229.374557.786.14020.8535.513464.953276812813289654.957596.256.74418.9861.702533.1565536128165664129.827504.798.34715.33138.174475.231310721281131200315.402415.5710.78011.87326.182402.232621441281262272910.215288.0015.5688.22925.783283.30
https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF#amd-ryzen-ai-max-395-strix-halo-128-gb-unified-memoryAMD Ryzen AI Max+ 395 (Strix Halo, 128 GB unified memory)
Step-3.7-flash-Q4_K_S
llama-batched-bench.exe -m Step-3.7-flash-Q4_K_S.gguf -c 65664 -b 2048 -ub 1024 -npp 0,2048,8192,16384,32768,65536 -ntg 128 -npl 1
PPTGBN_KVT_PP sS_PP t/sT_TG sS_TG t/sT sS t/s012811280.0000.004.87826.244.87826.242048128121769.367218.635.13424.9314.501150.0681921281832043.540188.155.50823.2449.048169.6316384128116512111.814146.535.94721.53117.761140.2232768128132896357.81991.586.77918.88364.59890.23655361281656641342.50148.828.49515.071350.99648.60 Step-3.7-flash-IQ4_XS
./llama-batched-bench -m Step-3.7-flash-IQ4_XS.gguf -c 65664 -b 2048 -ub 1024 -npp 0,2048,8192,16384,32768,65536 -ntg 128 -npl 1
PPTGBN_KVT_PP sS_PP t/sT_TG sS_TG t/sT sS t/s012811280.0000.005.93121.585.93121.582048128121768.143251.506.19420.6714.337151.7881921281832039.899205.326.52119.6346.420179.2316384128116512105.098155.896.89118.57111.989147.4432768128132896338.64596.767.79316.42346.43994.95655361281656641310.82050.009.48913.491320.30949.73 Step-3.7-flash-Q3_K_L
./llama-batched-bench -m Step-3.7-flash-Q3_K_L.gguf -c 262272 -b 2048 -ub 1024 -ctk q8_0 -ctv q8_0 -npp 0,2048,8192,16384,32768,65536,131072,262144 -ntg 128 -npl 1
PPTGBN_KVT_PP sS_PP t/sT_TG sS_TG t/sT sS t/s012811280.0000.005.01525.535.01525.5320481281217610.246199.885.07325.2315.319142.0481921281832037.229220.055.34123.9642.570195.441638412811651279.234206.785.48923.3284.723194.8932768128132896179.697182.355.81022.03185.507177.3365536128165664436.593150.116.57719.46443.169148.1713107212811312001262.377103.839.12414.031271.501103.1926214412812622723487.92175.1611.39111.243499.31274.95
https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF#5-acknowledgments5. Acknowledgments
This release stands on the work of the following authors and communities:
- bartowski— for
calibration\_datav5, the community-standard imatrix calibration anchor used by countless GGUF releases. Used for calibration purposes only; no license has been verified for this resource. - eaddario— for the
imatrix\-calibrationdataset (MIT), providing multilingual / code / math splits that form the backbone of this release’s domain balance - NousResearch— for
hermes\-function\-calling\-v1(Apache-2.0), used for agent / tool-call calibration coverage - ggml-org / llama.cpp— for the entire quantization and inference toolchain (MIT)
https://huggingface.co/stepfun-ai/Step-3.7-Flash-GGUF#6-license6. License
The GGUF quantization files in this repository are derivative works ofstepfun\-ai/Step\-3\.7\-Flashand are released under the sameApache 2.0license.
All calibration datasets retain their original licenses and are used strictly for quantization calibration purposes only.
Similar Articles
stepfun-ai/Step-3.7-Flash
Step 3.7 Flash is a 198B-parameter sparse MoE vision-language model with 11B active parameters per token, supporting 256k context and three reasoning levels, designed for high-throughput agentic workflows.
@AdinaYakup: Step-3.7-Flash New VL model from @StepFun_ai 198B / 11B active - MoE 256K context 3 reasoning level Up to 400 tokens/sec
StepFun releases Step-3.7-Flash, a new large vision-language MoE model with 198B parameters (11B active), 256K context, and up to 400 tokens/sec inference speed.
Stepfun 3.7 Flash is very good
Stepfun 3.7 Flash is a compact vision model that achieves aesthetics close to GLM 5.1 and 80% of its 3D world understanding, while using only 25% of the parameters, making it highly RAM-efficient.
Step 3.7 Flash IQ4_XS GGUF with preserve_thinking
Step 3.7 Flash is now available as an IQ4_XS GGUF quantized model, enabling efficient local inference with preserve_thinking support via llama.cpp, vLLM, Ollama, and other tools.
StepFun 3.7 Flash
StepFun released Step 3.7 Flash, a high-efficiency multimodal model optimized for real-world agentic tasks, featuring improved coding benchmarks (SWE-Bench Pro, Terminal-Bench) and compatibility with multiple agent harnesses.