KyleHessling1's picture
Initial upload: Jackrong DeepSeek-V4-Flash 9B eval with base 9B comparison
bcab855 verified
|
Raw
History Blame Contribute Delete
4.1 kB
metadata
title: Qwen3.5-9B-DeepSeek-V4-Flash · Q5 Evaluation
emoji: 🌿
colorFrom: green
colorTo: blue
sdk: static
pinned: false
license: mit
short_description: DeepSeek-V4 distill vs Qwen 9B base, head-to-head
tags:
  - evaluation
  - qwen3.5
  - deepseek
  - distill
  - gguf
  - llama.cpp
  - tool-calling
  - front-end

Qwen3.5-9B-DeepSeek-V4-Flash · Q5_K_M evaluation

A side-by-side benchmark of Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash (the DeepSeek-V4 distill) against Qwen/Qwen3.5-9B (the official post-trained 9B), both at Q5_K_M, on the same hardware and same eval suite. The fair-comparison control group answers the question: what does this distill actually add?

Headline

  • 0 / 5 agentic cap hits on the distill, 3 / 5 on the base — the base spirals on long-horizon thinking and hits the 8K token budget without producing a final answer; the distill commits and finishes
  • 2.2× faster agentic wall time on the distill (96.7 s vs 213.9 s, same 5 prompts)
  • 2.6× less reasoning chars on the distill across the agentic suite — tighter, more decisive thinking style
  • Distill clearly wins on creative front-end design — eye-review comparison shows the base produces visibly flatter output with animation and overlay bugs on the open-ended prompts (designer portfolio, mobile app marketing). On templated forms (dashboard, pricing) they're close. The distill ships output that punches above its parameter count.
  • 5 / 6 tool calls PASS on both — the distill doesn't break instruction-following or structured output
  • ~143 tok/s on both — same engine, same arch, same quant, same throughput. The distill is free in the literal sense.

Why this comparison

Reasoning distills usually carry a tax somewhere — instruction-following, throughput, output quality at the seams. The right way to look for that tax is to compare against the same base the distill came from, at the same parameter count, on the same hardware. This eval does that.

Credit: distill by Jackrong on Hugging Face. Base model by Qwen team. Eval by @KyleHessling1.

What's in this Space

  • index.html — card grid with every featured distill output linked, runtime stats up top
  • report.html — full writeup with side-by-side base vs distill tables for agentic, design, and tool-calling
  • design_*.html — five front-end design outputs (saas landing, analytics dashboard, designer portfolio, pricing, mobile marketing)
  • agentic_*.txt — five agentic reasoning outputs with thinking traces
  • *.thinking.txt — separated reasoning content for the design runs
  • *.raw.txt — raw model output before HTML extraction

The creative-coding (canvas/WebGL/Three.js) category was run on both models but not featured — most outputs in both had rendering bugs, which is an honest 9B-class weakness on shader/canvas math, not a distill question. Raw bench data is in the source repo.

Setup

Item Value
Distill Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash Q5_K_M (locally quantized, 6.1 GB)
Base Qwen/Qwen3.5-9B Q5_K_M from bartowski (6.4 GB)
Runtime llama.cpp cuda-12.8 (b8708), --flash-attn on, --jinja, embedded chat templates
Context 40,960 tokens, FP16 KV cache
Hardware RTX 5090 (32 GB), Intel Core Ultra 7 265K, 125 GB RAM
Thinking Always on for both

Numbers at a glance

Distill (Jackrong) Base (Qwen)
Agentic cap hits 0 / 5 3 / 5
Agentic wall time 96.7 s 213.9 s
Agentic completion tokens 13,806 31,142
Agentic reasoning chars 40,780 106,333
Design cap hits 0 / 5 1 / 5
Design wall time 376.0 s 493.8 s
Tool calling 5 PASS, 1 PARTIAL 5 PASS, 1 PARTIAL
Avg tok/s 141.9 145.5

License

MIT for this evaluation. Models are under their own respective licenses — see the linked model cards.