0xBakeer's TandemLLM recipe: Ornith-1.5-122B, EXL3 4.0 bpw

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 / 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 baseline0 / 1

Measured 2026-10-21 (browser check re-run 2026-10-22). Raw files: run info · command · final choice · contract · sampling · gate stats · gate log · screenshot · check · clip.mp4 · web.mp4 · web-poster.jpg.

Browser check 8.0 out of 10: the page loads, runs without console errors, is not blank, shows a world (85,242 parts, 1,056,446 triangles, English text), flies an automatic camera tour and animates; the 2 lost points are the frame rate (0.67 frames per second in software rendering without a GPU, 20 are needed) and the orbit check (mouse drag and wheel changed the picture, but not beyond what the page animates on its own). The check that ran together with the agent run stopped at its screenshot step (timeout 30 s). It was run again on the same machine from the saved files with the timeout raised to 120 s; the page file is the same one (sha256 9596e74c...). The agent finished by itself after 76.6 minutes, one try, no help; the “finish now” message (from 90 minutes) was not needed. 9 versions of the page were seen, 3 loaded without errors, and the last one counts. The limit per answer was 131,072 tokens, the value on the Ornith model card (README lines 175 and 187: max_new_tokens=131072). TandemLLM's own default is 32,768 tokens and applies only when a request sets none, so I raised the limit; the server setup is unchanged. My test gate shortened the limit of one request to what was left of the context (108,765 instead of 131,072). 37 requests, no error; the longest answer had 39,046 tokens and none was cut off (34 requests ended with a tool call, 3 with a normal stop). Thinking was on with the model's sampling (temperature 0.6, top_p 0.95, top_k 20) and the server default effort xhigh, the setting of 0xBakeer's own Ornith runs. A big flickering light is often in view and takes up a lot of room (my note on the world).

run_showcase.sh --harness opencode --task VX --upstream http://127.0.0.1:41631/v1 --model TandemLLM --sampling-file $RUN/sampling-show.json --out $RUN/raw/plan-2110-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 2026-10-22. Raw files: speed result · speed contract · log.

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 instruct (non-thinking) sampling of the model card (temperature 0.6, top_p 0.95, top_k 20, min_p 0). The author's own speed table (CHANGELOG.md lines 19 to 25) was measured greedy, thinking off, on prompts of 3,146 to 4,165 tokens and on TandemLLM commit 88b8070e, an earlier build than the 0.13.0 code (pack card README.md line 35); it lists 85.3 tok/s for prose and 114.2 tok/s for code, and it names sampled decoding as not measured (CHANGELOG.md line 27). Ours uses the card's sampling and our own prompts, so the draft acceptance and the speed are not directly comparable. TandemLLM's lookup corpus was not built (the author's corpus is not published); the lookup drafts from the conversation only, as after the README's two install commands. The headline 44.24 tok/s is the mean of the code prompt (43.85) and the prose prompt (44.63), medians of two runs each. With 2, 4 and 8 requests at once the total stays between 41.07 and 42.26 tok/s: the recipe runs one slot, so each request gets its share. Our test requests send cache_prompt false. TandemLLM does not read that field (it is not among the fields its engine reads, TandemLLM v0.13.0 docs/server.md, line 46), so its prefix cache stayed on, as in a normal setup; the memory test shows cached tokens from 16K up. The speed test starts every prompt with a unique line instead, and none of its 44 requests reported cached tokens. TandemLLM v0.14.0 came out on 22 Oct 2026, after this run. This page shows v0.13.0 and was not repeated. From its release notes: “Every prompt attention reads the KV heads as they are (SRV-673, SRV-681). At 131k one layer's NemotronH MTP step went from 28.5 to 1.6 ms, and a 179,678-token Ornith replay from 23.2 to 38.3 tok/s, with the same bits.”

python3 speed2.py --base-url http://127.0.0.1:39267/v1 --label run-speed-r0 --prompts $BENCH/repos/rigmark-v2/prompts-v2.json --output $RUN/raw/plan-2210-speed-r0/work/speed2 --contract $RUN/raw/plan-2210-speed-r0/work/speed2/speed-contract.json --parts decode,prefill,conc --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/ornith-1.5-122b-exl3-4.0/tokenizer.json --tokenizer-python $HOME/not published --sampling {"temperature": 0.6, "top_p": 0.95, "top_k": 20, "min_p": 0.0} --sampling-source https://github.com/0xBakeer/TandemLLM/blob/67ff8901a112f6477f97aff1748273991db124db/ops/profiles/_ornith.env lines 25-27 (TEMPERATURE 0.6, TOP_K 20, TOP_P 0.95, 'the checkpoint's generation_config.json') and https://huggingface.co/0xBakeer/TandemLLM-Ornith-1.5-122B-EXL3-4.0/blob/5330fb28b08ed1ff6d21c6723ed1b49bc06c2887/generation_config.json (temperature 0.6, top_k 20, top_p 0.95); docs/tool-eval-bench.md line 111 (the model's own sampling defaults are temperature 0.6, top_k 20, top_p 0.95); the model maker's card https://huggingface.co/ornith-ai/Ornith-1.5-122B/blob/e09bc0dfbe/README.md line 155 ('precise coding tasks': temperature 0.6, top_p 0.95, top_k 20, min_p 0, presence_penalty 0). min_p 0.0 from that card (no min_p in generation_config.json). No separate thinking set is documented: same values with thinking on/off. presence_penalty 0 = not sent. --model TandemLLM --no-cache-prompt --server-log $RUN/logs/server-A.log --engine-version-file $RUN/logs/engine-version.txt --server-slots 1 --server-slots-source 0xBakeer/TandemLLM@67ff8901 (v0.13.0) ops/profiles/_ornith.env:22 (QSE_PARALLEL=1; the shipped default ops/serve.env:322 is also QSE_PARALLEL=1); CHANGELOG.md:14-16 (one slot)

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 2026-10-22.

Note from the run: 2 of 400 MMLU-Pro answers were cut off at the 48-token limit, 5 of 40 LiveCodeBench answers at their limit of 32,768 tokens and 6 of 30 GPQA answers with thinking at the limit of 16,384 tokens; these count as wrong. Thinking was off for the LiveCodeBench test, yet 4 of the 40 replies contained a think block (the request flag was off in every request; cause not found). The 5 replies that ran into the limit are all long working-out text. GPQA Diamond ran on the public mirror fingertap/GPQA-Diamond (the original is gated). These are my own tests of this pack.

bash tools/quality2/measure_all.sh tools/quality2/cards/ornith15-122b-exl3-4.0bpw-tandemllm0130-1x.json --parts mmlupro --job plan-2110

bash tools/quality2/measure_all.sh tools/quality2/cards/ornith15-122b-exl3-4.0bpw-tandemllm0130-1x.json --parts gpqa --job plan-2110

bash tools/quality2/measure_all.sh tools/quality2/cards/ornith15-122b-exl3-4.0bpw-tandemllm0130-1x.json --parts ifeval --job plan-2110

bash tools/quality2/measure_all.sh tools/quality2/cards/ornith15-122b-exl3-4.0bpw-tandemllm0130-1x.json --parts lcb --job plan-2110

bash tools/quality2/measure_all.sh tools/quality2/cards/ornith15-122b-exl3-4.0bpw-tandemllm0130-1x.json --parts ruler --job plan-2110

bash tools/quality2/measure_all.sh tools/quality2/cards/ornith15-122b-exl3-4.0bpw-tandemllm0130-1x.json --parts ruler --job plan-2110

bash tools/quality2/measure_all.sh tools/quality2/cards/ornith15-122b-exl3-4.0bpw-tandemllm0130-1x.json --parts think --job plan-2110

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 2026-10-22. Raw files: result of every step · log · contract.

Note from the run: The largest size tested is 262,144 tokens, the setup's own context, so this is the largest size tested, not necessarily the largest the engine could load. The series ends at the first size that does not answer, so every size from 4K to 256K loaded and answered. The memory number is the rise over the idle value (unified memory) at the largest step and includes the model weights. Decode at the largest size is the server's own timing for the 64 tokens after the first one: 44.95 s, so 1.42 tok/s. Reading the 249,085-token prompt is not in that number; it took another 312.5 s (First token column, about 797 tok/s over the whole prompt). The prompt is a fixed filler text and the answer is short (65 tokens), so this is a memory test, not a speed test. TandemLLM v0.14.0 came out on 22 Oct 2026, after this run. This page shows v0.13.0 and was not repeated. From its release notes: “Every prompt attention reads the KV heads as they are (SRV-673, SRV-681). At 131k one layer's NemotronH MTP step went from 28.5 to 1.6 ms, and a 179,678-token Ornith replay from 23.2 to 38.3 tok/s, with the same bits.”

bash tools/quality2/measure_all.sh tools/quality2/cards/ornith15-122b-exl3-4.0bpw-tandemllm0130-1x.json --parts fit --job plan-2110

Also made this possible

• ModelOrnith AIHugging Face GitHub Website
• EXL3 formatturboderp@turboderp_ GitHub Hugging Face

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

Model

0xBakeer/TandemLLM-Ornith-1.5-122B-EXL3-4.0 @5330fb28 (weights byte-equal to the engine pin 92f7483c) (Ornith-1.5-122B by Ornith AI, Qwen3.5-MoE)

Engine

TandemLLM 0.13.0 on one DGX Spark, reasoning effort xhigh — the program that loads the model and answers requestsBuild and settings we used: TandemLLM 0.13.0 (0xBakeer/TandemLLM, StairCut, profile ornith-1.5-122b-exl3-4.0)

Recipe, engine (TandemLLM) and EXL3 4.0 bpw pack

github.com/0xBakeer/TandemLLM by 0xBakeer

Context window

262,144 tokens: how much text the model can see at once

Steps

1. Get the recipe from the list above. The list has no link for the engine (TandemLLM 0.13.0 on one DGX Spark, reasoning effort xhigh). The list has no direct download link for the model file itself. The base model is huggingface.co/ornith-ai/Ornith-1.5-122B.

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

   Copy

   ops/start.sh with ops/active-profile = ornith-1.5-122b-exl3-4.0 (profile ops/profiles/ornith-1.5-122b-exl3-4.0.env), run/local.env \
     REASONING_EFFORT=xhigh; engine :8000; requests via :8888 (adds X-Tandem: raw); via GB10 box 1-ornith-tandemllm-0130.sh start

TandemLLM 0.13.0 (0xBakeer/TandemLLM, StairCut, profile ornith-1.5-122b-exl3-4.0); ops/start.sh with ops/active-profile = ornith-1.5-122b-exl3-4.0 (profile ops/profiles/ornith-1.5-122b-exl3-4.0.env), run/local.env REASONING_EFFORT=xhigh; engine :8000; requests via :8888 (adds X-Tandem: raw); via GB10 box 1-ornith-tandemllm-0130.sh start

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 2026-10-22. 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.

python3 tools/quality2/sparkdash_client.py tools/quality2/cards/ornith15-122b-exl3-4.0bpw-tandemllm0130-1x.json <box> --client <client machine> --job plan-2110 --out $RUN

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