howtospark recipe: KAT-Coder V2.5-Dev (vLLM, 1 GB10 box)

What it builds as an agent

Task: Voxel world (3D scene, one HTML file). One attempt, no help. Counted: 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)0.5 / 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)1 / 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 2026-10-26. Raw files: check · run info · command · final choice · contract · sampling · screenshot · clip.mp4 · web.mp4.

The run's own headless check browser, which renders the page the model built, used up to 1352 % CPU. The load monitor counted it as another process. Browser check 8.5 out of 10: the page loads without a page error (ready after 0.01 s), has no console errors, is not blank and shows a world; the lost points are half a point on the frame rate (software rendering, no GPU) and one point on the automatic camera tour. The agent (OpenCode 1.19.3) ran 9 minutes 29 seconds, one try, no help, and stopped on its own, long before the 'finish now' mark at 90 minutes; its answer used 20,632 tokens. 5 versions of the page were seen on the way, 2 loaded without errors; the last clean one counts (index.html, sha256 1ab2ede0..., 5 snapshots). My rating of the world it built: 1.5 of 10 (rated in the browser, 27 Oct 2026). KAT-Coder is a coding model by Kwaipilot; this showcase does not measure its strength, and the page says so.

run_showcase.sh --harness opencode --task VX --upstream http://127.0.0.1:45595/v1 --model kat-coder --sampling-file $HOME/quality2-runs/kat-2610b/sampling-show.json --out $RUN/kat-2610b-show-r4/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 2026-10-26. Raw files: speed result · speed contract · log · long answer.

Requests at once: the server has 4 slots (from the author's recipe). C8 exceeds the slots: the extra requests wait in the queue, and they were measured anyway. The speed test ran with the author's setup: the howtospark recipe 'kat-coder-2-5-nvfp4' (recipe line: 'No patches, no custom flags'), vLLM 0.24.0 (torch 2.11.0+cu130), NVFP4 MoE via FLASHINFER_CUTLASS, MTP speculative decoding k=2 with a grafted Qwen3.6-35B-A3B MTP head (unsloth file @1c3f884b), context 262,144 tokens (recipe --max-model-len), KV pool 6 GiB (274,528 positions), 4 slots (--max-num-seqs 4), exactly as the recipe specifies. Two host fixes were needed to get it running, named openly: python3.12-dev headers are not in the recipe's requirements (fetched without sudo, no system change) and the FlashInfer JIT compiler started one process per core and hit a global out-of-memory (limited to MAX_JOBS=5; no vLLM flag changed). Setup from zero to the first answer: 27 minutes, 9 of them lost to that compiler OOM. The headline 64.57 tok/s is the median of the code prompt (64.57) and the prose prompt (64.57), two runs each (spread 1.6 % and 4.6 %), one warm-up excluded, seed 42, thinking off, no prompt cache. Drafts were accepted at a rate of 56.1 % (2.12 tokens per round). First token: 0.31 s at 2K, 1.09 s at 8K, 5.23 s at 32K prompt; prefill 6,270 tok/s at 32K. At the same time: 62.4 / 105.5 / 131.6 / 130.1 tok/s total at 1 / 2 / 4 / 8 requests. Eight requests share the recipe's 4 slots, so the total stops rising and each of the 8 gets 23.4 tok/s. This is the fastest single-Spark decode we have measured so far (next best in the class: 56.15 tok/s, Qwen3.8-Flash-Next EXL3 3.05 bpw, 25 Oct 2026).

python3 speed2.py --base-url http://127.0.0.1:35895/v1 --label kat-2610b-speed-r0 --prompts $BENCH/repos/rigmark-v2/prompts-v2.json --output $RUN/kat-2610b-speed-r0/kat-2610b-speed-r0/speed2 --contract $RUN/kat-2610b-speed-r0/kat-2610b-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/models/hf/KAT-Coder-V2.5-Dev-NVFP4-MTP/tokenizer.json --tokenizer-python $HOME/nachtplan-2510/venvs/vllm/bin/python --sampling {"temperature": 0.7, "top_p": 0.8, "top_k": 20, "presence_penalty": 1.5} --sampling-source Kwaipilot/KAT-Coder-V2.5-Dev @7be56fe7 README.md:319-324 (thinking: max_tokens 81920, temperature 1.0, top_p 0.95, presence_penalty 1.5, top_k 20) and README.md:346-352 (non-thinking: max_tokens 32768, temperature 0.7, top_p 0.8, presence_penalty 1.5, top_k 20, chat_template_kwargs enable_thinking false). min_p not named -> not sent. --model kat-coder --no-cache-prompt --server-log $RUN/server-A.log --engine-version-file $RUN/engine-version.txt --server-slots 4 --server-slots-source howtospark recipe kat-coder-v2-5-nvfp4 step 4 (--max-num-seqs 4) = start script flint-2510/kat-ctl.sh; KV pool 6 GiB = 1.05x at the full 262,144 window (recipe 'Key vLLM parameters': --kv-cache-memory-bytes). The recipe's value is the largest it documents.

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 2026-10-26.

Note from the run: Quality ran on this card on 26 Oct 2026, all seven tests of the fixed method: MMLU-Pro 400, GPQA Diamond 198, IFEval 200, LiveCodeBench 40 (code runs), long texts 8K/32K, Ruler 8K/32K, plus GPQA 30 with thinking. Thinking off in all tests except the thinking test (limit 16,384 tokens). Sampling is the maker's (instruct: temperature 0.7, top_p 0.8, top_k 20, from the model README), seed 42, the questions ran 4 at a time, the recipe's slot count. A cut-off answer counts as wrong: 10 of the 40 LiveCodeBench answers were cut at their limit, and 107 answers across the quality tests in total. The error margin sits next to every number. These are my own tests; no other run is mixed into the numbers on this page.

bash tools/quality2/measure_all.sh tools/quality2/cards/kevin-katcoder-v25dev-35b-nvfp4-mtpgraft-vllm0240-1x.json --parts mmlupro --job kat-2610b

bash tools/quality2/measure_all.sh tools/quality2/cards/kevin-katcoder-v25dev-35b-nvfp4-mtpgraft-vllm0240-1x.json --parts gpqa --job kat-2610b

bash tools/quality2/measure_all.sh tools/quality2/cards/kevin-katcoder-v25dev-35b-nvfp4-mtpgraft-vllm0240-1x.json --parts ifeval --job kat-2610b

bash tools/quality2/measure_all.sh tools/quality2/cards/kevin-katcoder-v25dev-35b-nvfp4-mtpgraft-vllm0240-1x.json --parts lcb --job kat-2610b

bash tools/quality2/measure_all.sh tools/quality2/cards/kevin-katcoder-v25dev-35b-nvfp4-mtpgraft-vllm0240-1x.json --parts ruler --job kat-2610b

bash tools/quality2/measure_all.sh tools/quality2/cards/kevin-katcoder-v25dev-35b-nvfp4-mtpgraft-vllm0240-1x.json --parts ruler --job kat-2610b

bash tools/quality2/measure_all.sh tools/quality2/cards/kevin-katcoder-v25dev-35b-nvfp4-mtpgraft-vllm0240-1x.json --parts think --job kat-2610b

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 rise of used memory (MemTotal - MemAvailable) over the idle value before the step, on the one box. Unified memory, so GPU and CPU share it. Measured 2026-10-26. Raw files: result of every step · log · contract.

Note from the run: The memory test climbed the fixed ladder up to the recipe's full window: the model answered at 262,144 tokens of context (the recipe's --max-model-len), so nothing larger was tested. The peak rise over idle was 42.24 GiB at the largest step (the number includes the model weights). Decode speed at these sizes is not measured.

bash tools/quality2/measure_all.sh tools/quality2/cards/kevin-katcoder-v25dev-35b-nvfp4-mtpgraft-vllm0240-1x.json --parts fit --job kat-2610b

Also made this possible

• Model (KAT-Coder V2.5-Dev)KwaipilotHugging Face GitHub
• Grafted MTP head (Qwen3.6-35B-A3B-NVFP4-Fast @1c3f884b)Unsloth@unslothai Hugging Face GitHub Website
• Engine (vLLM 0.24.0)vLLM projectGitHub

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 DGX Spark class box (GB10, 'kevin'), 128 GB unified memory, standing GPU clock cap 2200 MHz (the default application clock of the GB10 is 2418 MHz; the cap is the standing setting on these boxes and the run used it). The model needs about 21.9 GB of NVFP4 weights; the box also runs two small standing services (they were idle-checked before the run).

Model

KAT-Coder V2.5-Dev 35B-A3B NVFP4 (sakamakismile/KAT-Coder-V2.5-Dev-NVFP4 @be2e4d17; base Kwaipilot/KAT-Coder-V2.5-Dev @7be56fe7) + Qwen3.6-35B-A3B MTP head graft (unsloth/Qwen3.6-35B-A3B-NVFP4-Fast @1c3f884b)

Model (KAT-Coder V2.5-Dev)

huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev

NVFP4 build with grafted MTP head and recipe (howtospark 'kat-coder-2-5-nvfp4')

huggingface.co/sakamakismile/KAT-Coder-V2.5-Dev-NVFP4

Grafted MTP head (Qwen3.6-35B-A3B-NVFP4-Fast @1c3f884b)

huggingface.co/unsloth/Qwen3.6-35B-A3B-NVFP4-Fast

Engine

Kwaipilot's coding model (35B-A3B) as a community NVFP4 build with a grafted MTP head, vLLM 0.24.0, 262,144-token context, 4 slots — the program that loads the model and answers requests · github.com/vllm-project/vllmBuild and settings we used: vLLM 0.24.0 (torch 2.11.0+cu130), FLASHINFER_CUTLASS NVFP4 MoE, MTP speculative k=2 (howtospark recipe)

Steps

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

2. Start the server with this command. It loads the model into memory.

   Copy

   systemd-run \
     --user \
     --unit=kat-serve ... vllm serve KAT-Coder-V2.5-Dev-NVFP4-MTP \
     --served-model-name kat-coder \
     --max-model-len 262144 \
     --kv-cache-memory-bytes 6442450944 \
     --gpu-memory-utilization 0.85 \
     --max-num-seqs 4 \
     --max-num-batched-tokens 8192 \
     --speculative-config mtp k=2 + maker parser flags

   The file names are the ones we used. Run the command in the folder that holds them, or put the full path in front of each name.

vLLM 0.24.0 (torch 2.11.0+cu130), FLASHINFER_CUTLASS NVFP4 MoE, MTP speculative k=2 (howtospark recipe); systemd-run --user --unit=kat-serve ... vllm serve $HOME/models/hf/KAT-Coder-V2.5-Dev-NVFP4-MTP --served-model-name kat-coder --max-model-len 262144 --kv-cache-memory-bytes 6442450944 --gpu-memory-utilization 0.85 --max-num-seqs 4 --max-num-batched-tokens 8192 --speculative-config mtp k=2 + maker parser flags (flint-2510/kat-ctl.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 2026-10-26. 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 2.50.0 (@21a3511) ran its own tests on 27 Oct 2026 (job kat-2610c-sd): its QA set 82.67 % (3 of 150 answers cut) and its MMLU set 89.82 % (5 cut), temperature 0, thinking off, a cut-off answer counts as wrong. Its own column, never mixed into Speed v2, Quality v2 or the ranking. The sparkDash client ran on another machine and was never part of the measured group.

python3 tools/quality2/sparkdash_client.py tools/quality2/cards/kevin-katcoder-v25dev-35b-nvfp4-mtpgraft-vllm0240-1x.json <box> --client <client machine> --job kat-2610c-sd --out $RUN

tool-eval-bench --base-url http://127.0.0.1:<port> --model kat-coder --short --no-think --timeout 120 --json-file $BENCH/sparkdash-client/runs/kat-2610c-sd-sparkdash-r0/tooleval/result.json --no-live