Cheblan/ThinkingCap-Qwen3.8-Flash-Next-strixw

ThinkingCap-Qwen3.8-Flash-Next (Strixite .strixw format for AMD Strix Halo)

This repository contains the converted and quantized weights of bottlecapai/ThinkingCap-Qwen3.8-Flash-Next in the native .strixw format for the high-performance Strixite inference engine on AMD Strix Halo APUs (Ryzen AI Max+ 395 / gfx1151).

---

⚡ Highlights

• Architecture: 180B Total Parameters:
  • 125B MoE Backbone: 48 layers (36 Gated Delta Net recurrent layers + 12 Full Attention layers), 512 fine-grained experts (10 active per token, ~6B active parameters).
  • 51B PLE N-Gram Table (ngram.table): 3-gram pre-trained lookup embeddings (20M base vocabulary) for high-speed recurrent indexing.
  • 4B MTP (Multi-Token Prediction): Speculative multi-token drafting (7 drafts ahead).
  • Multimodal Vision Tower (vision.strixw): Full vision & image input capability.
• Format: Optimized .strixw layout with mixed Q4/Q8 quantization:
  • MoE Experts: exp_gu=q4g128, exp_down=q4g128, sh_gu=q8g64, sh_down=q8g64
  • Recurrent GDN: gdn_in=q4g128, gdn_out=q8g64
  • Attention QKV & LM Head: attn_qkv=q8g64, attn_o=q8g64, lm_head=q8g64
  • Embeddings: BF16
• Context Capacity: Up to 262,144 tokens with full static K/V pre-commit support.
• Speed: 50–70+ tokens/sec decode throughput on AMD Ryzen AI Max+ 395 (128 GB LPDDR5X-8533 unified memory).

---

📦 Files in this Repository

File	Size	Description
weights.strixw	~66 GB	Main quantized model weights (MoE + GDN + Attention)
ngram.table	~49 GB	51B parameter PLE (Pre-trained Lookup Embeddings) n-gram table
vision.strixw	~857 MB	Multimodal vision tower
tokenizer.json	~13 MB	Fast tokenizer
config.json	~5 KB	Model configuration & architecture metadata
generation_config.json	~200 B	Recommended sampling parameters (temperature 0.3, top_p 0.85)
chat_template.jinja	~9 KB	Qwen chat template

---

🖥️ System Requirements

• Hardware: AMD Ryzen AI Max+ 395 / Ryzen AI Max 390 (Strix Halo, gfx1151, 40 CUs RDNA 3.5).
• RAM: 128 GB LPDDR5X (Unified Memory Architecture).
• Runtime: ROCm 6.3+ / Linux kernel 6.12+.
• Inference Engine: Strixite (strix_server).

---

🚀 Running with Strixite

1. Configuration (strix-server-thinkingcap.conf)

# --- Model Files ---
model-id = thinkingcap-qwen3.8-flash-next
weights = /path/to/weights.strixw
ngram = /path/to/ngram.table
vision = on
tokenizer = /path/to/tokenizer.json
generation-config = /path/to/generation_config.json
prompt-cache-dir = /path/to/prompt-cache

# --- Serving & Concurrency ---
host = 0.0.0.0
port = 5300
total-slots = 3

# --- Memory & Context ---
capacity = 262144
total-capacity = 786432
kv-precommit = on
chunk = 16384
chunk-shared = 2048

# --- Precision & Kernels ---
gdn-precision = half-u
attention-precision = fast
kv-cache = q8

# --- Multi-Token Prediction (MTP) ---
mtp = on
mtp-draft = 7
mtp-margin = 1.5

# --- Prompt Cache (RAM + NVMe Tier) ---
low-memory = adapt
prompt-cache-gib = 128
prompt-cache-ram-margin-gib = 8
prompt-cache-idle-s = 60
prompt-cache-write-gib-per-hour = 48
prompt-cache-checkpoints = 32
prompt-cache-delta-max-tokens = 65536

2. Launching the Server

strix_server --config strix-server-thinkingcap.conf

---

🏷️ Attribution & Base Model

• Original Base Weights: bottlecapai/ThinkingCap-Qwen3.8-Flash-Next by BottleCap AI.
• Underlying Architecture: Alibaba Cloud Qwen team (Qwen/Qwen3.8-Flash-Next).
• ThinkingCap is post-trained to reduce internal reasoning tokens by 38.5% on average while retaining ~88.8% macro accuracy, delivering faster, more cost-effective inference for agentic coding and reasoning tasks.

---

📜 License

This model inherits and is distributed under the Qwen Community License 1.0 (full terms provided in LICENSE).

• Free for research, personal, and commercial applications under 100M MAU and $20M monthly revenue.
• Commercial Model-as-a-Service (MaaS) hosting and primary AI-assisted coding/work assistants require an agreement from Alibaba/Qwen.

---

📚 Citation

If you use this model or these weights, please cite the original BottleCap AI team:

@misc{ThinkingCap-Qwen3.8-Flash-Next,
  title     = {bottlecapai/ThinkingCap-Qwen3.8-Flash-Next},
  author    = {Osusky, Adam and Lindauer, Jan and Jirkovsky, Adam and Mihal, Filip and Platek, Ondrej and Herel, David and Ihnatchenko, Luka and Bartek, Vojtech and Jirak, Jiri and Kubista, Daniel and Krus, Frantisek and Mikolov, Tomas},
  year      = {2026},
}