New: Products & Members. Sell courses, downloads and subscriptions straight from your website.Take a look
Qwen (Alibaba) · 125B parameters, 6B active

Qwen3.8 Flash-Next

UD-Q4_K_XL GGUF · llama.cpp (b10840) · tested on 1 Oct 2026

123.6tokens/s at 1k
109.3tokens/s at 32k
0.79s to first token
4/5coding passed
7.0/7agentic points
5/5puzzles right
2/2German met
93.6GB VRAM used

VerdictAI

Strong for coding and agent work on a single workstation GPU. On hard tasks, though, it thinks for minutes, and long prompts take noticeably long to the first token.

Strengths

  • About 124 tokens/s on one card
  • 4 of 5 one-shot coding tasks work
  • Agent task 3 of 3, puzzles 5 of 5
  • Clean, natural German

Weaknesses

  • Thinks for minutes on hard tasks
  • Ran out of tokens once (3D scene)
  • 22.6 s to the first token at 32k context
  • Four long prompts in parallel: over a minute to the first token

Speed

256 output tokens per request, measured with one and with four concurrent requests.

Output per request in tokens/s1 request4 concurrent
1k context
1 request123.6
4 concurrent57.0
8k context
1 request115.0
4 concurrent30.6
32k context
1 request109.3
4 concurrent15.3
Time to first token in secondslower is better1 request4 concurrent
1k context
1 request0.79
4 concurrent3.05
8k context
1 request5.72
4 concurrent16.23
32k context
1 request22.57
4 concurrent66.91
ContextconcurrentFirst token (s)Prefill (tokens/s)Output per request (tokens/s)Output total (tokens/s)VRAM max (GB)
1k10.791,370123.6124.193.6
1k43.051,41357.0217.593.6
8k15.721,541115.0115.593.6
8k416.231,55130.654.593.6
32k122.571,540109.3109.893.6
32k466.911,26215.311.493.7

Output total: all generated tokens divided by the time from the first to the last token across all concurrent requests.

One-shot coding

Each task gets one attempt and a budget of 81,920 tokens for thinking and answer. The suite ran 2 tasks at the same time, so the speed applies under that load. You can try the apps right here. They run sealed off on demo.techatlab.com, without network or storage, and only start after a click. As with any web page, their code is visible in the browser.

Tetris-style game

passed
Still from the recording: Tetris-style game
Generated by Qwen3.8 Flash-Next (AI)Watch clip

Keyboard controls: click into the game first.

Token budget used41,912 of 81,920 tokens

440 s · 95.4 tokens/s

Fully playable with ghost piece, next-piece preview, score and levels. The game-over message appears below the board instead of on top of it. AI

Balls in a spinning hexagon

passed
Still from the recording: Balls in a spinning hexagon
Generated by Qwen3.8 Flash-Next (AI)Watch clip
Token budget used60,509 of 81,920 tokens

704 s · 86.0 tokens/s

The hexagon spins, all 20 balls bounce off each other and off the walls, none escapes. The balls are rather small. AI

3D scene in a canvas

failedcut off
Still from the recording: 3D scene in a canvas
Generated by Qwen3.8 Flash-Next (AI)Watch clip

No runnable file, so only the recording.

Token budget used81,920 of 81,920 tokens

1,057 s · 77.8 tokens/s

Stopped at the 81,920-token budget while the code was still being written. The file ends mid-text and the page stays empty. AI

SVG illustration

passed
Still from the recording: SVG illustration
Generated by Qwen3.8 Flash-Next (AI)Watch clip
Token budget used68,160 of 81,920 tokens

837 s · 81.8 tokens/s

Moonlit cabin with falling snow, twinkling stars and a moving aurora, plus a title nobody asked for. The chimney smoke is barely visible. AI

Small web app (finance tracker)

passed
Still from the recording: Small web app (finance tracker)
Generated by Qwen3.8 Flash-Next (AI)Watch clip
Token budget used41,243 of 81,920 tokens

422 s · 99.6 tokens/s

Complete dashboard with search, filters, donut and cash-flow charts and export. The tallest cash-flow bars run past the top. AI

Agentic

A customer case with 4 simulated tools and 7 checks, 3 runs.

CheckRun 1Run 2Run 3
Found the customer by the right email addressyesyesyes
Checked the shipment of the latest orderyesyesyes
Issued the credit exactly once, to the right customeryesyesyes
Correct credit amount (€37.26)yesyesyes
Sent an email to the customeryesyesyes
Email states amount and shipment statusyesyesyes
No errors and no needless callsyesyesyes
7 of 74 tool calls · 9.6 s7 of 74 tool calls · 6.3 s7 of 74 tool calls · 5.9 s

Puzzles

5 of 5 puzzles solved correctly.

Puzzleexpectedanswerresulttokens
Counting letters99right417
Age puzzle3636right713
Divisibility20002000right983
Knights and knavesknaveknaveright407
Domino tiling8989right404

German

Two writing tasks in German with fixed requirements.

TaskWordsSentencesRequirements
Customer email (German)818met
AI explained for kids (German)775met

Prompts

The tasks word for word, exactly as the model received them.

One-shot coding

Tetris-style game
Create a polished Tetris-style falling-block game. Use a standard 10x20 board, all 7 tetrominoes with distinct colors, rotation with simple wall kicks, soft drop (down arrow), hard drop (space), line clears with a short flash animation, score, level and a next-piece preview, increasing speed per level, and a game-over screen with restart on R. Controls: left/right arrows move, up arrow rotates. The game starts immediately.

Requirements for your answer: output ONE complete, self-contained HTML file in a single ```html code block. Inline all CSS and JavaScript. Do not use any external libraries, CDNs, web fonts, images or network requests; it must work offline by simply opening the file in Chrome. It must fill a 1920x1080 browser window nicely and start immediately on load without requiring a click.
Balls in a spinning hexagon
Write a physics simulation of 20 numbered balls bouncing inside a spinning hexagon. The hexagon rotates continuously around its center. The balls are affected by gravity and friction, collide realistically with the rotating walls (take the wall's motion into account) and with each other. Every ball has its own color and shows its number. Implement collision detection and response yourself, no physics library.

Requirements for your answer: output ONE complete, self-contained HTML file in a single ```html code block. Inline all CSS and JavaScript. Do not use any external libraries, CDNs, web fonts, images or network requests; it must work offline by simply opening the file in Chrome. It must fill a 1920x1080 browser window nicely and start immediately on load without requiring a click.
3D scene in a canvas
Render an animated 3D scene using only the HTML canvas (raw WebGL, or Canvas 2D with your own projection math; no libraries): a small low-poly island with trees and a lighthouse whose light beam rotates, surrounded by animated water, while the camera slowly orbits the island. Include simple directional lighting and correct depth ordering.

Requirements for your answer: output ONE complete, self-contained HTML file in a single ```html code block. Inline all CSS and JavaScript. Do not use any external libraries, CDNs, web fonts, images or network requests; it must work offline by simply opening the file in Chrome. It must fill a 1920x1080 browser window nicely and start immediately on load without requiring a click.
SVG illustration
Create a detailed SVG illustration of a cozy mountain cabin at night: snowy mountains, a pine forest, a cabin with warm glowing windows and smoke from the chimney, a starry sky with a moon and green northern lights. Use gradients and layering for depth. Add subtle animation (twinkling stars, drifting smoke, waving aurora) with SVG or CSS animation. The SVG must scale to fill the whole window.

Requirements for your answer: output ONE complete, self-contained HTML file in a single ```html code block. Inline all CSS and JavaScript. Do not use any external libraries, CDNs, web fonts, images or network requests; it must work offline by simply opening the file in Chrome. It must fill a 1920x1080 browser window nicely and start immediately on load without requiring a click.
Small web app (finance tracker)
Build a small personal finance tracker web app: add income and expense entries (description, amount, category, date) via a form, show a list that can be sorted and filtered, live totals and the balance, and a donut or bar chart of expenses per category drawn with canvas or SVG (no chart library). Persist the data in localStorage and pre-fill 12 realistic example entries on first load so the app looks populated. Modern, clean design.

Requirements for your answer: output ONE complete, self-contained HTML file in a single ```html code block. Inline all CSS and JavaScript. Do not use any external libraries, CDNs, web fonts, images or network requests; it must work offline by simply opening the file in Chrome. It must fill a 1920x1080 browser window nicely and start immediately on load without requiring a click.

Agentic

System
You are an operations assistant for an online shop. Today is 2026-09-30. Use the available tools to complete the task step by step. Do not invent data; look it up. When everything is done, reply with a short summary for the human operator.
User
Customer anna.berg@example.com is asking where her most recent order is. Check its shipping status. Our policy: if a shipment is more than 5 days past its promised delivery date, issue a store credit of 15% of that order's total (rounded to cents), then email the customer a short update with the shipment status and the credit amount. If it is not late, just email the status.
Tools
lookup_customer, get_shipment, issue_credit, send_email

Puzzles

Counting letters
How many times does the letter "r" appear in the phrase "strawberry raspberry cranberry"?

Think it through, then end your reply with a final line in the form: ANSWER: <answer>
Age puzzle
Alice is twice as old as Bob was when Alice was as old as Bob is now. Their ages add up to 63. How old is Alice?

Think it through, then end your reply with a final line in the form: ANSWER: <answer>
Divisibility
What is the sum of all integers n with 1 <= n <= 100 such that n^2 + 1 is divisible by 5?

Think it through, then end your reply with a final line in the form: ANSWER: <answer>
Knights and knaves
On an island, knights always tell the truth and knaves always lie. A says: "B is a knave." B says: "A and C are of the same type." Is C a knight or a knave? Answer with one word.

Think it through, then end your reply with a final line in the form: ANSWER: <answer>
Domino tiling
In how many different ways can a 2 x 10 rectangle be completely tiled with 1 x 2 dominoes?

Think it through, then end your reply with a final line in the form: ANSWER: <answer>

German

Customer email (German)
Schreibe eine kurze, freundliche E-Mail (hoechstens 120 Woerter) an einen Kunden, dessen Paket sich um eine Woche verspaetet. Entschuldige dich, erklaere kurz den Grund (Engpass beim Versanddienstleister) und biete einen Gutschein ueber 10 Prozent an. Sie-Form.
AI explained for kids (German)
Erklaere einem zwoelfjaehrigen Kind in genau fuenf Saetzen, wie ein KI-Sprachmodell funktioniert. Nutze eine Alltagsanalogie und vermeide Fachbegriffe.

Settings

Parameters
125B parameters, 6B active
Quantization
UD-Q4_K_XL GGUF: Unsloth GGUF, speculative decoding draft from the model's own MTP head (Q8_0)
Runtime
llama.cpp (b10840)
Context window
262,144 tokens
Sampling
temperature 1.0, top_p 0.95, top_k 20, min_p 0.05
Hardware
RTX PRO 6000 Blackwell (96 GB), Intel Core i9-14900K, 64 GB RAM
Suite
v1.0.0
Run
1 Oct 2026, 03:43 to 04:24
License
Qwen Community 1.0
Released
26 Aug 2026

Techat's production model, measured directly on the running production server at low night-time load. VRAM figures include the KV cache reserved for 262k context. The coding part ran with a budget of 81,920 tokens after Tetris had stopped while still thinking at 32k tokens. AI

Who runs the tests

Techat Lab is a project by Luca Hansmann. Luca also runs Techat, a platform with the AI assistant Techie, and this page promotes it. The measurements are independent of that: same suite, same machine.

The test machine

GPU
NVIDIA RTX PRO 6000 Blackwell, 96 GB VRAM
CPU
Intel Core i9-14900K
RAM
64 GB
Runtime
llama.cpp (llama-server)

AI notice AI

The test suite collects the measurements automatically. Coding verdicts, conclusions and short texts are written with the help of Techie, Techat's AI assistant, from measurements and recordings. Such texts carry the “AI” mark.

How we test

techat

Use these models in Techat

Techie, the AI assistant in Techat, normally runs on Qwen3.8 Flash-Next from this lab. It handles your website, email, bookings and contacts right in the chat.

Start for freeWhat Techat doesNo credit card, cancel any month