Mia's TP4 recipe: DeepSeek-V4.1-Flash-1017

What it builds as an agent

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

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 2026-10-20. Raw files: check · run info · command · final choice · contract · sampling · 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 1408 % CPU. The load monitor counted it as another process. Browser check 9.0 out of 10: the page loads, runs without console errors and is not blank; the only lost point is the frame rate, measured at 1.62 frames per second in software rendering without a GPU. The server's per-answer cap of 32,768 tokens was lifted to the card's 262,144 (named on this page). The agent ended by itself after 90.8 minutes with finish reason stop; the “finish now” message was sent at 90 minutes.

run_showcase.sh --harness dsh-021 --task VX --upstream http://127.0.0.1:39895/v1 --model deepseek-v4.1-flash --sampling-file $RUN/sampling-show.json --out $RUN/raw/plan-2010-show-r3/show --limit-s 7200 --inject-after-s 5400 --poll-s 20 --ctx 1048576

How fast it runs

Method: code+prose; 2 run(s) each; 512 tokens; 1 warm-up(s); median. Measured 2026-10-20. Raw files: speed result · speed contract · long answer.

The code prompt is not steady: its two runs gave 35.7 and 65.1 tok/s (median 50.4, spread 41 %), the prose prompt 51.8 and 52.1 tok/s (median 52.0). The headline 51.19 tok/s is the mean of the two prompt medians. A measurement above 15 % spread is repeated: this is the repeat, and the first measurement (replaced by this one) showed the same pattern (spread 40 %), so the spread repeats, but I did not find the cause. Prefill at 8K runs at two speeds: its three runs gave 2,582, 2,745 and 4,969 tok/s (the median 2,745 is the slower speed), and the first measurement had the same two speeds.

python3 speed2.py --base-url http://127.0.0.1:35689/v1 --label run-speed-r7 --prompts $BENCH/repos/rigmark-v2/prompts-v2.json --output $RUN/raw/plan-2010-speed-r7/work/speed2 --contract $RUN/raw/plan-2010-speed-r7/work/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 --tokenizer-json $HOME/NewModels/DeepSeek-V4.1-Flash-1017/tokenizer.json --tokenizer-python $HOME/not published --sampling {"temperature": 1.0, "top_p": 1.0} --sampling-source Model card first (Niklas Oct 20 00:2x, E1): deepseek-ai/DeepSeek-V4.1-Flash-1017 README.md:77-84 recommends temperature 1.0, top_p 1.0; Mia's TP4 repo MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks@64b02dea docs/tp4.md:42-44 used top_p 0.95 (older card state), not adopted. The TP4 repo names no top_k, so top_k is not sent. Same setting for thinking and non-thinking. Speed v2 runs with thinking off (enable_thinking false per request), show/think with thinking on. Always send max_tokens. --model deepseek-v4.1-flash --no-cache-prompt --server-log $RUN/logs/server-A.log --engine-version-file $RUN/logs/engine-version.txt

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 2026-10-20.

Note from the run: 57 of 400 MMLU-Pro answers and 55 of 198 GPQA Diamond answers were cut off at the 48-token limit (the model starts with “Let's go step by step” instead of the letter) and count as wrong. Without them, 257 of the 343 finished MMLU-Pro answers are right (74.9 %) and 69 of the 143 finished GPQA Diamond answers (48.3 %). The headline numbers keep the cut-off answers as wrong, as the method says. 2 of the 80 long-text answers were cut off at their limit of 128 tokens and count as wrong too.

bash tools/quality2/measure_all.sh tools/quality2/cards/ds1017-sglang-tp4-4x.json --parts mmlupro --job plan-2010

bash tools/quality2/measure_all.sh tools/quality2/cards/ds1017-sglang-tp4-4x.json --parts gpqa --job plan-2010

bash tools/quality2/measure_all.sh tools/quality2/cards/ds1017-sglang-tp4-4x.json --parts ifeval --job plan-2010

bash tools/quality2/measure_all.sh tools/quality2/cards/ds1017-sglang-tp4-4x.json --parts lcb --job plan-2010

bash tools/quality2/measure_all.sh tools/quality2/cards/ds1017-sglang-tp4-4x.json --parts ruler --job plan-2010

bash tools/quality2/measure_all.sh tools/quality2/cards/ds1017-sglang-tp4-4x.json --parts ruler --job plan-2010

bash tools/quality2/measure_all.sh tools/quality2/cards/ds1017-sglang-tp4-4x.json --parts think --job plan-2010

How much context fits

Memory test. The recipe fixes the context (1,048,576 tokens), so the server is started once with the recipe's own command and reserves its memory up front. Then, for each size (4K, 8K, 16K ... up to the largest size tested), one prompt that fills about 95 % of that size is sent, and the model must answer. The largest size that answered counts; it is the highest size tested, not a limit of the setup. Decode speed at that size was not measured. Peak memory is, on each of the four boxes, the rise of used memory (MemTotal - MemAvailable) over the idle value before the step; the table and the number show the highest of the four boxes. Unified memory, so GPU and CPU share it. Measured 2026-10-20. Raw files: result of every step · log · contract.

Note from the run: Highest TESTED step is 131072 tokens (steps 4K to 128K all loaded and answered, no failure); the context in the recipe is 1048576, so 131072 is not the limit. Peak 100.84 GiB = highest node (box-1) at the largest step, without the load phase; with the load phase up to 103.8 GiB per node (box-3 103.8, box-4 103.0, box-1 102.7, box-2 101.7). Decode speed at the steps not measured. The memory is read on all four boxes; the number shown is the highest of them.

bash tools/quality2/measure_all.sh tools/quality2/cards/ds1017-sglang-tp4-4x.json --parts fit --job plan-2010

Also made this possible

• ModelDeepSeekHugging Face @deepseek_ai GitHub Website
• Engine (SGLang)SGLang (sgl-project)GitHub Website

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

4x NVIDIA GB10 boxes, tensor parallel 4 over a switchless CX7 ring: 2x ASUS Ascent GX10, 1x Gigabyte AI TOP ATOM, 1x Lenovo ThinkStation PGX. One box (an ASUS GX10) runs an older driver (580.173.02) and kernel (6.17.0-1031-nvidia) than the other three (580.178.04, 7.0.0-1019-nvidia). That is named openly.

Model

DeepSeek-V4.1-Flash-1017 (official checkpoint)

Engine

SGLang, tensor parallel on 4 GB10 boxes — the program that loads the model and answers requests · github.com/sgl-project/sglangBuild and settings we used: SGLang (lmsysorg/sglang:dev-dsv41 pin 3dbc3130) + Mia TP4 production line (canary-roce, no mesh) @64b02dea, DSpark k=5, TP4 ring

Recipe (TP4 production line)

github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks by Mia (MiaAI-Lab)

Steps

1. Get the recipe and the engine from the list above. The list has no direct download link for the model file itself. The base model is huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017.

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

   SERVED_MODEL_NAME=deepseek-v4.1-flash \
     PORT=8888 ./start-tp4-switchless.sh serve

   What the parts mean

   • PORT=8888the network port the server listens on

SGLang (lmsysorg/sglang:dev-dsv41 pin 3dbc3130) + Mia TP4 production line (canary-roce, no mesh) @64b02dea, DSpark k=5, TP4 ring; SERVED_MODEL_NAME=deepseek-v4.1-flash PORT=8888 ./start-tp4-switchless.sh serve ($HOME/dsv41-4x-spark on box-3; ranks box-3=0 box-4=1 box-1=2 box-2=3; ENV_FILE .env.tp4-switchless = production line without mesh)

Mia's own benchmark, run unchanged

This is Mia's own benchmark tool (commit 21a3511), run unchanged. It has its own section here and its own column on the class 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 2026-10-20. 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) on a client machine (never in the measured group), Quality: seed 1234, temperature 0, thinking off, a truncated answer counts as wrong; decode, prefill and tool eval send no seed; own row, never mixed with quality-v2.

python3 tools/quality2/sparkdash_client.py tools/quality2/cards/ds1017-sglang-tp4-4x.json <box> --client <client machine> --job plan-2010 --out $RUN

tool-eval-bench --base-url http://127.0.0.1:<port> --model deepseek-v4.1-flash --short --no-think --timeout 120 --json-file $BENCH/sparkdash-client/runs/work/tooleval/result.json --no-live