BobClawblaw's TensorFold recipe: Qwen3.8-27B (MLX 4-bit + DFlash2)

What it builds as an agent

Task: Airship dock (3D scene, one HTML file). One attempt, no help. Counted: the final file counts.

Design and build the most impressive, creative, elaborate and detailed steampunk airship dock voxel world you are capable of. Theme: a steampunk airship dock. Everything else (style, story, extra details) is your choice.

How you work:
- You work alone in the folder /work. Nobody will answer questions. Aim to finish in about 60 minutes. After 90 minutes you will get a message to finish; then you have at most 30 more minutes. Hard stop after 120 minutes. Spend the time where it shows in the result.
- Tools on this machine: shell, files, python3 with numpy and Pillow, ffmpeg, node, and `shot.sh index.html` (opens the page in a headless browser, saves a screenshot and prints console errors). There is no internet and no access to other machines.
- You may create your own assets locally with these tools (textures, sky, sounds, music) and embed them in the file as data URIs. Do not download anything.
- Keep a working /work/index.html at all times: write a first working version early, then improve it step by step. If you are stopped, the last version that loaded without errors counts.
- Check your work with `shot.sh` as often as you like, fix every error, and keep improving until you are satisfied. Then stop.

Deliverable: ONE self-contained file /work/index.html that opens and runs in Chrome.
- JavaScript libraries: only these, through an import map with these exact URLs (they work offline in the test browser). Use whatever helps, including all Three.js addons (post-processing, sky, water, noise and so on):
  - "three": "https://cdn.jsdelivr.net/npm/three@0.170.0/build/three.module.js"
  - "three/addons/": "https://cdn.jsdelivr.net/npm/three@0.170.0/examples/jsm/"
  - "simplex-noise": "https://cdn.jsdelivr.net/npm/simplex-noise@4.0.3/dist/esm/simplex-noise.js"
  - "@tweenjs/tween.js": "https://cdn.jsdelivr.net/npm/@tweenjs/tween.js@25.0.0/dist/tween.esm.js"
  - "cannon-es": "https://cdn.jsdelivr.net/npm/cannon-es@0.20.0/dist/cannon-es.js"
- No other external resources. Build every object from cubes (voxels); InstancedMesh is allowed.
- The page fills the whole window and resizes with it. All visible text is in English.
- After the scene is built, set `window.scene` to the Three.js scene and `window.ready = true`.
- An automatic slow cinematic camera tour starts on load. Mouse input switches to OrbitControls (drag to rotate, scroll to zoom).
- Sound, if any, starts only after the first click (browsers block autoplay).
- A small label in the top-left corner: "Steampunk Airship Dock".

Minimum content (go far beyond it):
1. A tall iron docking tower with platforms, stairs or ladders, and a mooring arm at the top.
2. At least two airships, each with a balloon envelope, a gondola and propellers.
3. Big brass gears on the tower or the buildings.
4. Pipes and at least three chimneys that emit rising steam particles.
5. At least one crane and a stack of cargo crates.
6. A ground area with a workshop building and a small rail track.
7. Propellers and gears rotate, steam rises and fades, and one airship slowly approaches the tower, docks, waits and leaves again, in a loop.
8. Warm sunset lighting with shadows.

The browser check

• page loads without page error1 / 1
• window.ready === true within 30 s1 / 1
• no console errors (whole session)1 / 1
• frame is not blank1 / 1
• frame rate in software GL (rAF over 5 s)0 / 1
• required objects (voxel mass, >=3 steam emitters, >=2 airships, label)1.5 / 2
• motion: pixel diff 0.5, >=3 rotating nodes 0.5, large group translating 12 / 2
• OrbitControls: drag and wheel change the picture beyond the animation baseline0 / 1

Measured 2026-10-25. Raw files: check · run info · command · final choice · contract · sampling · gate stats · screenshot · clip.mp4 · web.mp4 · web-poster.jpg.

The run's own headless check browser, which renders the page the model built, used up to 1589 % CPU. The load monitor counted it as another process. What I noticed in the browser: a big airship flies in from the side and docks at the mooring tower, and it is easy to miss, because it only docks after about 30 seconds. Browser check 7.5 out of 10 (airship-dock task): the page loads without a page error, is ready after 0.08 s, has no console errors, is not blank, and the scene moves (the motion check gave its 2 points). The 2.5 lost points are the frame rate (1.6 frames per second in software rendering without a GPU, 10 give half a point and 20 the full point), the orbit-controls test (drag and wheel changed the picture less than the check needs) and half a point of the objects check: it found no steam source where 3 are needed, although the finished page draws steam plumes (see the video above). The agent (Qwen Code 0.25.2) finished by itself after 31.1 minutes, one try, no help; the “finish now” message (from 90 minutes) was not needed. 13 versions of the page were seen on the way, all 13 loaded without errors, and the last one counts. 66 requests, no error, no request blocked (9,939,512 prompt tokens, 178,161 answer tokens); 64 ended with a tool call and 2 with a normal stop, none at the limit. The last 4 requests are the harness's own memory step after the build. The largest prompt had 222,608 tokens. The limit per answer was 262,144 tokens, the value on the Qwen3.8-27B model card (README lines 507 to 510); my test proxy lowered it to the room left in the 262,144-token context in 42 of the 66 requests. The proxy estimates the size of a prompt without the server's tokenizer, and the estimate was up to 1.72 times the real size (see gate-stats.json); the lowest limit it set was 19,503 tokens and no answer reached its limit. That is a property of my test tool, not of the setup. Thinking was on with the model card's sampling (temperature 1.0, top_p 0.95, top_k 20), the context was 262,144 tokens, 1 stream.

run_showcase.sh --harness qwen-code --task K1 --upstream http://127.0.0.1:45059/v1 --model qwen38-27b --sampling-file $RUN/sampling-show.json --out $RUN/raw/nacht2410-show-r4/show --limit-s 7200 --inject-after-s 5400 --poll-s 20 --ctx 262144

Extra build: Voxel world (not part of the score)

The 32 GB class is ranked on the airship-dock task. This voxel-world run used a different prompt (the one the 128 GB+ classes use), so it is shown as an extra and does not count toward the score or the ranking.

Task: Voxel world (3D scene, one HTML file). One attempt, no help. Which file is used: the final file counts.

Design and create the most impressive, creative, elaborate and detailed
voxel art world you are capable of. The theme and style are ENTIRELY your
choice — pick whatever world you can execute best and push it as far as
you can: scale, variety, detail, color, animation, atmosphere, life.
Surprise me.

Hard requirements:
- Everything ends up in ONE single self-contained HTML file (game.html in
  the current working directory) that opens and runs directly in Chrome.
  Use WebGL and whatever libraries you like, but embed everything —
  no internet access at view time, no external files.
- The scene must present itself: include an automatic slow cinematic
  camera tour (orbiting the world, occasionally zooming in on highlights)
  that starts on load, plus mouse controls (orbit/zoom) for manual viewing.
- All user-facing text (UI, HUD, titles, descriptions, hints) must be in
  ENGLISH, regardless of the language of any project files you encounter.

When finished, verify your HTML actually loads and renders without
console errors, then state DONE with the absolute path of game.html.
Do not ask questions.

How you work (same for every model):
- You work alone in the folder /work (this is the current working directory). Nobody will answer questions. Aim to finish in about 60 minutes. After 90 minutes you will get a message to finish; then you have at most 30 more minutes. Hard stop after 120 minutes.
- Tools on this machine: shell, files, python3 with numpy and Pillow, ffmpeg, node, and `shot.sh game.html` (opens the page in a headless browser, saves a screenshot and prints console errors). There is no internet and no access to other machines.
- Keep a working /work/game.html at all times. If you are stopped, the last version that loaded without errors counts.
- JavaScript libraries available offline in the test browser, through an import map with these exact URLs: "three": "https://cdn.jsdelivr.net/npm/three@0.170.0/build/three.module.js", "three/addons/": "https://cdn.jsdelivr.net/npm/three@0.170.0/examples/jsm/", "simplex-noise": "https://cdn.jsdelivr.net/npm/simplex-noise@4.0.3/dist/esm/simplex-noise.js", "@tweenjs/tween.js": "https://cdn.jsdelivr.net/npm/@tweenjs/tween.js@25.0.0/dist/tween.esm.js", "cannon-es": "https://cdn.jsdelivr.net/npm/cannon-es@0.20.0/dist/cannon-es.js". No other external resources.
- After the scene is built, set `window.scene` to the Three.js scene (if you use Three.js) and `window.ready = true`.

The browser check

• page loads without page error1 / 1
• window.ready === true within 30 s1 / 1
• no console errors (whole session)1 / 1
• frame is not blank1 / 1
• frame rate in software GL (rAF over 5 s)0 / 1
• world: scene size 0.5 (parts >= 80 or >= 20000 triangles; canvas fallback), visible text English if any 0.51 / 1
• automatic camera tour: the view changes on its own in two successive ~5 s windows (1 each)2 / 2
• animation in the world (nodes / instances / particles moving besides the camera)1 / 1
• OrbitControls: drag and wheel change the picture beyond the animation baseline1 / 1

Measured 2026-10-25. Raw files: check · run info · command · final choice · contract · sampling · gate stats · screenshot · clip.mp4.

The run's own headless check browser, which renders the page the model built, used up to 1569 % CPU. The load monitor counted it as another process. Extra build, not part of the score. This run used the voxel-world prompt (the one the 128 GB and larger classes use), not the airship-dock prompt this page is ranked on. Same box, same server settings, same agent as the airship-dock run above; only the prompt differs. Browser check 9.0 out of 10: the page loads without a page error, is ready after 0.01 s, has no console errors, is not blank, shows a world (15,874 parts, 192,208 triangles, English text), flies an automatic camera tour in both windows (70 % and 54 % of the picture changed), animates (43 animated nodes) and answers to mouse drag and wheel. The 1 lost point is the frame rate (7.56 frames per second in software rendering without a GPU, 10 give half a point and 20 the full point). My view of the world: relatively little detail, but what is there looks round and smooth, without bigger bugs. The agent (Qwen Code 0.25.2) finished by itself after 33.5 minutes, one try, no help; the “finish now” message (from 90 minutes) was not needed. 10 versions of the page were seen on the way, 5 of them loaded without errors (the others had a declared-twice or used-before-defined error), and the last one counts. 98 requests, no error, no request blocked (12,286,064 prompt tokens, 211,985 answer tokens); 95 ended with a tool call and 3 with a normal stop, none at the limit. The last 4 requests are the harness's own memory step after the build. The largest prompt had 226,004 tokens. The limit per answer was 262,144 tokens, the value on the Qwen3.8-27B model card; my test proxy lowered it to the room left in the 262,144-token context in 48 of the 98 requests (the lowest limit it set was 15,305 tokens; no answer reached its limit). The proxy's estimate of a prompt size was up to 1.37 times the real size (see gate-stats.json). Thinking was on with the model card's sampling (temperature 1.0, top_p 0.95, top_k 20), 1 stream.

run_showcase.sh --harness qwen-code --task VX --upstream http://127.0.0.1:44861/v1 --model qwen38-27b --sampling-file $RUN/sampling-show.json --out $RUN/raw/nacht2410-show-r0/show --limit-s 7200 --inject-after-s 5400 --poll-s 20 --ctx 262144

How fast it runs

Method: code+prose; 2 run(s) each; 512 tokens; 1 warm-up(s); median; requests at once: C1,C2,C4,C8 at once; 2 rounds each; 512 tokens per request; prose prompt; total = all tokens / wall time, per request = tokens / end-to-end time; median over the rounds. Measured 2026-10-25. Raw files: speed result · speed contract · log · long answer.

Requests at once: the server has 1 slot (from the author's recipe). C2, C4, C8 exceed the slots: the extra requests wait in the queue, and they were measured anyway. The speed test ran with the author's setup: BobClawblaw's recipe (BobClawblaw/qwen38-27b-vision-rtx-5090, commit 2147ce9d), served with his serve.sh unchanged (only host and port added). Engine: TensorFold 1.1.2 at his branch q35-prompt-int8 (BobClawblaw/TensorFold, commit 18d93451), built here with the recipe's setup.sh (zig 0.17.0, CUDA 13.3); the weights and git sources came from our depot at the pinned revisions instead of being downloaded, everything else is the recipe. Weights: the MLX 4-bit Qwen3.8-27B of TensorFold (TensorFold/Qwen3.8-27B-MLX-4bit, revision 22d8d538) with z-lab's DFlash2 drafts (z-lab/Qwen3.8-27B-DFlash2, revision 50307d4c), an FP8 KV cache, int8 prompt matmuls, vision loaded, a context of 262,144 tokens and the recipe's --parallel auto (serve.sh:20-30). Sampling is the model card's: temperature 0.7, top_p 0.8, top_k 20, presence penalty 1.5 with thinking off; seed 42; no prompt cache in the requests (every request carries its own nonce). The headline 253.44 tok/s is the mean of the code prompt (268.17) and the prose prompt (238.71), medians of two runs each (spread 2.3 % and 3.2 %). Requests at once: the recipe serves one stream at the full 262,144 tokens (README.md:277: PARALLEL auto means as many streams at once as the window allows, one at 262,144). As my test method asks, I used the largest setting the server allows at the full context, which is 1 stream. The test sent 2, 4 and 8 requests at once anyway; the extra requests wait in the queue, so the total stays flat (205.5, 200.6, 205.5 and 204.6 tok/s for 1, 2, 4 and 8 at once) and the speed per request falls (205.5, 154.1, 85.0 and 46.5 tok/s). The decode values above are measured with one request; the 1-request value of the concurrency test (205.5 tok/s) counts the whole request, including the wait for the first token.

python3 speed2.py --base-url http://127.0.0.1:34391/v1 --label nacht2410-speed-r0 --prompts $BENCH/repos/rigmark-v2/prompts-v2.json --output $RUN/raw/nacht2410-speed-r0/nacht2410-speed-r0/speed2 --contract $RUN/raw/nacht2410-speed-r0/nacht2410-speed-r0/speed2/speed-contract.json --parts decode,prefill,conc,l1 --prefill-depths 2048,8192,32768 --prompt-ids code,prose --runs 2 --max-tokens 512 --warmups 1 --prefill-runs 3 --conc-levels 1,2,4,8 --conc-rounds 2 --conc-max-tokens 512 --conc-prompt prose --tokenizer-json $HOME/engines/bob-2410/models/Qwen3.8-27B-MLX-4bit/tokenizer.json --tokenizer-python $HOME/hfvenv/bin/python --sampling {"temperature": 0.7, "top_p": 0.8, "top_k": 20, "min_p": 0.0, "presence_penalty": 1.5} --sampling-source https://huggingface.co/Qwen/Qwen3.8-27B README.md lines 252 (Thinking Mode) and 253 (Instruct/non-thinking mode), local copy not published line 252-253 (not published) --model qwen38-27b --no-cache-prompt --server-log $RUN/logs/server-A.log --engine-version-file $RUN/logs/engine-version.txt --server-slots 1 --server-slots-source BobClawblaw/qwen38-27b-vision-rtx-5090@2147ce9d README.md:277 (PARALLEL auto = streams at once as many as the window allows: one at 262,144; more streams were not measured on this engine); serve.sh:28 (--parallel auto)

How well it answers

Thinking off, except the last test. Seed 42. Sampling from the model card. A cut-off answer counts as wrong. Measured 2026-10-25.

Note from the run: MMLU-Pro and GPQA Diamond ran with an answer limit of 48 tokens (my usual limit for these two tests, the answer is one letter); none of the answers of these two tests was cut off, and 1 of the 198 GPQA answers had no answer letter and counts as wrong. 1 of 40 LiveCodeBench answers was cut off at its limit of 32,768 tokens and 10 of 30 GPQA answers with thinking at the limit of 16,384 tokens; these count as wrong. None of the long-text answers was cut off. The questions of each test ran 1 at a time, the stream count of the recipe's server (same questions, limits and seed 42 for every question). Sampling: with thinking off temperature 0.7, top_p 0.8, top_k 20, presence penalty 1.5; with thinking on (GPQA 30) temperature 1.0, top_p 0.95, top_k 20 (the model card). These are my own tests, and no other run of this setup is mixed into the numbers on this page.

bash tools/quality2/measure_all.sh tools/quality2/cards/bob-tensorfold-1.1.2-q35-int8-mlx4bit-dflash2-5090.json --parts mmlupro --job nacht2410

bash tools/quality2/measure_all.sh tools/quality2/cards/bob-tensorfold-1.1.2-q35-int8-mlx4bit-dflash2-5090.json --parts gpqa --job nacht2410

bash tools/quality2/measure_all.sh tools/quality2/cards/bob-tensorfold-1.1.2-q35-int8-mlx4bit-dflash2-5090.json --parts ifeval --job nacht2410

bash tools/quality2/measure_all.sh tools/quality2/cards/bob-tensorfold-1.1.2-q35-int8-mlx4bit-dflash2-5090.json --parts lcb --job nacht2410

bash tools/quality2/measure_all.sh tools/quality2/cards/bob-tensorfold-1.1.2-q35-int8-mlx4bit-dflash2-5090.json --parts ruler --job nacht2410

bash tools/quality2/measure_all.sh tools/quality2/cards/bob-tensorfold-1.1.2-q35-int8-mlx4bit-dflash2-5090.json --parts ruler --job nacht2410

bash tools/quality2/measure_all.sh tools/quality2/cards/bob-tensorfold-1.1.2-q35-int8-mlx4bit-dflash2-5090.json --parts think --job nacht2410

How much context fits

Memory test. For each context size (4K, 8K, 16K ... up to the largest size tested for the setup) the server is started again with that context, one prompt that fills about 95 % of the context is sent, and the model must answer. The largest size that loaded and answered counts. Decode speed is measured on that answer at the largest size. Peak memory is the GPU memory in use (nvidia-smi); the GPU was empty before each start. Measured 2026-10-25. Raw files: result of every step · log · contract.

Note from the run: The context is the author's own 262,144 tokens, which is the largest size tested here; the sizes from 4K to 128K came before it. For each size the server was started again with that context (the CONTEXT setting of serve.sh), one prompt that fills about 95 % of the size was sent, and the model answered at every size. The memory number is the GPU memory in use on the 5090 at the largest size (27.33 GiB at 262,144 tokens, with a prompt of 249,036 tokens). The highest value over all sizes was 29.21 GiB, at 65,536 tokens. Decode speed at these sizes is not measured.

bash tools/quality2/measure_all.sh tools/quality2/cards/bob-tensorfold-1.1.2-q35-int8-mlx4bit-dflash2-5090.json --parts fit --job nacht2410

Also made this possible

• ModelQwenHugging Face @Alibaba_Qwen GitHub Website
• Engine (TensorFold)Ash Hart@ashxhart GitHub Hugging Face Website
• Model files (MLX 4-bit, revision 22d8d538)TensorFoldHugging Face @ashxhart GitHub Website
• Draft model (DFlash2, revision 50307d4c)z-lab (Zhijian Liu)@zhijianliu_ Hugging Face GitHub

Run it yourself

What you need and how to start it, step by step. The results below come from exactly this setup.

What you need

Hardware

1x NVIDIA GeForce RTX 5090 32 GB, one GPU.

Model

TensorFold/Qwen3.8-27B-MLX-4bit @22d8d538 + z-lab/Qwen3.8-27B-DFlash2 @50307d4c

Model files (MLX 4-bit, revision 22d8d538)

huggingface.co/TensorFold/Qwen3.8-27B-MLX-4bit

Draft model (DFlash2, revision 50307d4c)

huggingface.co/z-lab/Qwen3.8-27B-DFlash2

Engine

TensorFold engine (his branch q35-prompt-int8) for one RTX 5090, MLX 4-bit weights, DFlash2 drafts, FP8 KV cache, 262,144 tokens of context, one stream — the program that loads the model and answers requests · github.com/ashhart/TensorFoldBuild and settings we used: TensorFold 1.1.2 q35-prompt-int8 @18d93451 (BobClawblaw fork, zig 0.17.0, CUDA 13.3, sm_120), DFlash2 drafts, int8 prompts, vision on

Recipe and TensorFold branch (q35-prompt-int8)

github.com/BobClawblaw/qwen38-27b-vision-rtx-5090 by BobClawblaw

Steps

1. Get the recipe, the engine and the model files from the list above.

2. Start the server with this command. It loads the model into memory. When the model has loaded, the server waits for requests on port 8888.

   Copy

   PORT=8888 HOST=127.0.0.1 ./serve.sh

   What the parts mean

   • PORT=8888the network port the server listens on

TensorFold 1.1.2 q35-prompt-int8 @18d93451 (BobClawblaw fork, zig 0.17.0, CUDA 13.3, sm_120), DFlash2 drafts, int8 prompts, vision on; PORT=8888 HOST=127.0.0.1 ./serve.sh

Mia's own benchmark, run unchanged

This is Mia's own benchmark tool (commit 21a3511), run unchanged. It has its own section on this page, and it is never mixed with my numbers. A cut-off answer counts as wrong here, so a value can differ from sparkDash's own screen.

Seed 1234 applies to the two quality tests; decode, prefill and tool eval send no seed. The benchmark client ran on a separate machine that is never part of the measured group. Measured 2026-10-25. Version 2.50.0, commit 21a3511.

QA. Quality QA, 150 items, seed 1234, temperature 0, thinking off, 4 requests at once (Mia's default). An answer cut off at the token limit counts as wrong. ± is the 95 % margin (Wilson).

MMLU. Quality MMLU, 285 items (5 for each of 57 subjects, 4 answers), seed 1234, temperature 0, thinking off. An answer cut off at the token limit counts as wrong. ± is the 95 % margin (Wilson).

Decode. Decode bench, 1 request at once, 400 tokens, prompt type structured, temperature 0, thinking off. No seed is sent.

Prefill. Prefill bench at 4K, 8K, 16K and 32K tokens. sparkDash skips a size that does not fit into the context. No seed is sent.

Tool Eval Quick. Tool Eval Quick: tool-eval-bench (SeraphimSerapis, commit c61319b), 15 core scenarios (--short), --no-think, --timeout 120. Score 0 to 100. No seed is sent.

Note from the run: sparkDash runs its own tests (its own QA and MMLU questions at temperature 0, thinking off, a cut-off answer counts as wrong; none of the QA answers and 20 of the 285 MMLU answers were cut off), so its decode and prefill values are not comparable with Speed v2 above: its decode test uses a structured prompt, Speed v2 a code and a prose prompt. The sparkDash client ran on another machine and was never part of the measured group.

python3 tools/quality2/sparkdash_client.py tools/quality2/cards/bob-tensorfold-1.1.2-q35-int8-mlx4bit-dflash2-5090.json <box> --client <client machine> --job nacht2410 --out $RUN

tool-eval-bench --base-url http://127.0.0.1:<port> --model qwen38-27b --short --no-think --timeout 120 --json-file $BENCH/sparkdash-client/runs/nacht2410-sparkdash-r5/tooleval/result.json --no-live