README.md

metadata

title: Smaug-Flash Explorer
emoji: 🐉
colorFrom: indigo
colorTo: pink
sdk: static
pinned: false
license: mit
short_description: Specs, benchmarks, real encoder, BYO-endpoint chat

🐉 Smaug-Flash Explorer

An interactive companion page for abacusai/Smaug-Flash — Abacus.AI's agentic-coding fine-tune of deepseek-ai/DeepSeek-V4-Flash-0731 (304.18 B params, 166.9 GB checkpoint across 48 shards, 1 M context, MIT).

Everything runs client-side in your browser: no backend, no GPU, no quota.

Tab	What it actually does
🧭 Overview	Spec table from config.json and the Hub's safetensors metadata, checkpoint anatomy, training approach (2× SFT + 1× KTO LoRA merged as full attention-only deltas), known behaviors and inference gotchas — plus live Hub stats fetched from the public API.
📊 Benchmarks	The model card's numbers as Plotly charts: grouped Smaug-Flash vs base bars, a delta chart, and a LiveBench category radar — with the card's own figures embedded for comparison and its caveats spelled out.
🔤 Prompt encoder	The repo ships no Jinja chat template, so this tab runs the model's own encoding/encoding_dsv4.py — vendored here byte-for-byte unmodified and executed by Pyodide — to turn any message list into the exact prompt string the model was trained on. It also parses a raw completion back into OpenAI-format fields.
💬 Playground	A streaming chat client that speaks OpenAI-compatible SSE directly from your browser to an endpoint you choose: your own vLLM/SGLang deployment, a tunnelled remote one, or the HF router (for a plumbing check against a live Flash-class model).
🚀 Deploy & size	The card's vLLM and SGLang recipes, the reference inference/ conversion path, a Python client snippet, and an interactive weight-memory calculator that maps quantization choices onto real GPU tiers and prices.

Why there is no in-Space inference

The checkpoint is 304.18 B parameters / 166.9 GB on disk (packed-FP4 experts + block-FP8 attention) and its Hub inferenceProviderMapping is empty — no serverless Inference Provider serves it. HF Spaces hardware tops out at ZeroGPU's 96 GB, and the native-format budget is ~165 GB of weights before any KV cache, so the weights cannot be loaded in a Space. This page therefore does the things that do work at zero GPU cost.

To serve the model for real, see the Deploy & size tab — a dedicated multi-GPU tier for this model starts around $5.50/h. If you want a Space to serve it, a community GPU grant is the route.

Notes worth knowing

• Special-token delimiter. The repo's encoding_dsv4.py and its tokenizer.json delimit special tokens with U+FF5C (＼) rather than ASCII | — near-identical in most fonts, which is why a hand-typed pipe silently fails to parse. The decoder repairs typed ASCII pipes and tells you when it did.
• reasoning_effort ∈ low (0-char preamble) / high (476 chars) / max (526 chars); None is treated as low, it only applies in thinking mode, and only on the first message. Budget up to 384 K output tokens for high/max.
• The encoder is strict by design: malformed completions raise (missing thinking, missing EOS token, Unexpected content at end) instead of being patched up — the card is explicit about that.
• Benchmark values in the charts are digitized from the model card's figures; the originals are embedded on the page and remain authoritative.

Running it locally

git clone https://huggingface.co/spaces/xdLocutus/Smaug-Flash-Explorer
cd Smaug-Flash-Explorer
python -m http.server 8765      # open http://127.0.0.1:8765

Locally is also the easiest way to use the Playground against a local runtime (http://127.0.0.1:8000/v1): no TLS, no CORS friction.

Optional: the Gradio variant

app.py + requirements.txt in this repo are a Gradio port of the same five tabs (same data, same vendored encoder, plus an API/MCP surface). They are inert while the README frontmatter says sdk: static. To switch, replace the frontmatter with:

sdk: gradio
sdk_version: 6.31.0
app_file: app.py
python_version: "3.12"

That requires hardware — Gradio Spaces on a free account need PRO, and ZeroGPU additionally needs a 30-day-old account or an approved community grant. The Gradio app is designed for ZeroGPU: it keeps a no-op @spaces.GPU function so the runtime is attached but no quota is spent, since it never touches CUDA.

Attribution

Model card and repository contents (MIT) belong to Abacus.AI; encoding_dsv4.py is vendored unmodified from the model repo; the base architecture is DeepSeek-AI's. This Space is unofficial.