Step-3.5-Flash-PRISM

A "role-play" following unrestricted/unchained PRISM-LITE version of StepFun's Step 3.5 Flash intended particularly for over-refusal and propaganda mechanisms suppression using our SOTA PRISM pipeline.

PRISM-PRO version avialable here: hhttps://ko-fi.com/s/d70e27c5b5

For Full Custom Production PRISM versions & raw tensors reach out @ https://ko-fi.com/ex0bit.

☕ Support Our Work

If you enjoy our work and find it useful, please consider sponsoring or supporting us!

Option	Description
PRISM VIP Membership	Access to all PRISM models
Bitcoin	bc1qarq2pyn4psjpcxzp2ghgwaq6y2h4e53q232x8r

---

Model Highlights

• PRISM Ablation — State-of-the-art technique that removes over-refusal behaviors while preserving model capabilities
• 196B MoE Architecture — 196 billion total parameters with only 11 billion active per token across 288 fine-grained routed experts + 1 shared expert
• Multi-Token Prediction (MTP-3) — Predicts 4 tokens simultaneously, achieving 100–300 tok/s typical throughput (peaking at 350 tok/s)
• 256K Context Window — Cost-efficient long context via 3:1 Sliding Window Attention (SWA) ratio
• Frontier Reasoning & Coding — 97.3 on AIME 2025, 74.4% on SWE-bench Verified, 51.0% on Terminal-Bench 2.0
• Accessible Local Deployment — Runs on high-end consumer hardware (Mac Studio M4 Max, NVIDIA DGX Spark)

Model Architecture

Specification	Value
Architecture	Sparse Mixture-of-Experts (MoE)
Backbone	45-layer Transformer (4,096 hidden dim)
Total Parameters	196.81B (196B Backbone + 0.81B Head)
Activated Parameters	~11B (per token)
Routed Experts per Layer	288
Shared Experts	1 (always active)
Selected Experts per Token	Top-8
Vocabulary Size	128,896
Context Length	256K
Attention	Hybrid SWA (3:1 SWA-to-Full ratio)
MTP Head	Sliding-window attention + dense FFN (4 tokens/pass)

Benchmarks

Benchmark	Step 3.5 Flash	DeepSeek V3.2	Kimi K2.5	GLM-4.7	MiniMax M2.1
Agent
τ²-Bench	88.2	80.3	85.4	87.4	86.6
BrowseComp	51.6	51.4	60.6	52.0	47.4
GAIA (no file)	84.5	75.1	75.9	61.9	64.3
xbench-DeepSearch (2025.05)	83.7	78.0	76.7	72.0	68.7
Reasoning
AIME 2025	97.3	93.1	96.1	95.7	83.0
HMMT 2025 (Feb.)	98.4	92.5	95.4	97.1	71.0
IMOAnswerBench	85.4	78.3	81.8	82.0	60.4
Coding
LiveCodeBench-V6	86.4	83.3	85.0	84.9	—
SWE-bench Verified	74.4	73.1	76.8	73.8	74.0
Terminal-Bench 2.0	51.0	46.4	50.8	41.0	47.9

llama.cpp (GGUF)

For local deployment (requires ~120 GB VRAM for int4, smaller quants are available):

./llama-cli -m step3.5_flash_prism_Q4_K_S.gguf --jinja

Recommended Parameters

Use Case	Temperature	Top-P	Max New Tokens
Reasoning / Coding	1.0	0.95	32768
General Chat	0.6	0.95	4096

Hardware Requirements

Setup	Details
BF16 (Full)	8x H100/A100 80GB with tensor parallelism
FP8 Quantized	8x A100 80GB with expert parallelism
GGUF INT4 (Local)	~120 GB unified memory (Mac Studio M4 Max 128GB, DGX Spark, AMD Ryzen AI Max+ 395)

License

This model is released under the PRISM Research License.

Acknowledgments

Based on Step 3.5 Flash by StepFun AI. See the technical report and blog post for more details on the base model.