Tern-35B-A3B (matthewagi, Qwen3.6 35B-A3B MoE decision model, author server, self-hosted)v1.7 · 10 s rule

Composite (secondary)

32.1 · Composite rank #77

Cost and measurement conditions

$0.085 per 1,000 decisions (estimated)

ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 a8a05890ddab2d39d1a1bd543f2b5c83842bdc70aa54766788bc89b7a8a448d7): base string Qwen/Qwen3.6-35B-A3B (the card's base_model, confirmed by the author's export.json) priced as qwen/qwen3.6-35b-a3b USD 0.15/M input, USD 1.0/M output. Estimated, not charged, not a quote from the author or any host. Tokens as reported by the author server's own usage block (input_tokens = len(input_ids) of the encoded request: state plus all schema and template tokens; output_tokens = 0, one forward pass with a decision head, nothing generated). The offline count through a first-party transcription of the author's encoder and the pinned tokenizer predicted the server's numbers EXACTLY, item by item: 842,725 input tokens over the 1,485 answered items, and 15 HTTP 422 refusals at the pre-registered ids and positions; the 15 refused items count 0 tokens. OVER THE OFFICIAL COST CAP, DISCLOSED: USD 0.085124 per 1,000 decisions on the common basis (USD 0.084273 over all 1,500 items) is 1.318x the official cap USD 0.064595; the row is ranked (v1.6.16 vllm-sr-d3 and v1.6.24 imajev-9b precedent) and falls outside the official-caps board views by construction. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them); their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence.

x2 + 0.15 s (assumption, not measured)

Evaluator-owned rented Verda GPU VM, one NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB (driver 580.178.04), FIN-03 (Finland, EU), pay-as-you-go non-spot, instance type 1RTXPRO6000.30V at USD 2.525/h. The author's own `python3 -m tern serve` from the pinned commit on 127.0.0.1:8000 (vLLM backend on 127.0.0.1:8101) inside a network-none Docker container with the cut proven from inside (loopback only, no default route, DNS and TCP fail), read-only root, tmpfs scratch, read-only weights, harness and task input, empty credential placeholders, HF offline flags, one request at a time (no concurrency, no batching endpoint). Timed over loopback on the VM. One disclosed deviation: a read-write compile-cache bind shared between the synthetic smoke container and the scored container, so the second cold start fits the time limit; it carries compiler artifacts only and leaves content on the pod disk, which is deleted. Measured 22 Oct 2026 (VM reserved 06:29Z, removed 06:48Z) on the frozen v1.6-regular-20261012 self-hosted input (sha256 dfd36720..., checked locally, on the node and inside the container); determinism 50/50 bit-identical; node shredded and VM destroyed afterwards (teardown confirmed).

Model and serving disclosure

The author's own `tern` package (Matthew-agi/Tern at d154fa085d92eb2170681cd90817e18c735553ff, tree 529c3f40c3cc966b15b3e964a5c79ea9bde6c894, Apache-2.0, 66 files, every git blob oid verified; no tag exists on that repo; the README-tested commit 9723b745 differs only in README/docs), unmodified, installed from the pinned tree into our own image and started as `python3 -m tern serve --profile rtx-pro-6000 --model <pinned verified snapshot> --cache-dir /cache/tern --gpu 0 --host 127.0.0.1 --port 8000 --backend-port 8101` (the author's Docker command plus our local --model). Author profile kept unchanged: max_model_len 32,768, 32,768 batch tokens, 256 seqs, CUDA-graph capture 16,384, BF16, Triton recurrent kernels, head batch 16/16,384, CUBLAS_WORKSPACE_CONFIG=:4096:8, CLEF_SUBMISSION_MODE=single, all 40 layers and 8 experts, no result cache; the h100, h100-bf16 and a100 profiles and the HF runtime/serve.py were not used. Static source review PASS WITH FINDINGS, no blockers (no non-loopback network, no telemetry, no item logging, no pickle/trust_remote_code, no rule keyed to item text or benchmark names). DISCLOSED CODE PATTERNS, cleared by the benchmark owner (21 Oct 2026; author code on a rented Verda pod only, weights pre-staged, network none): (1) the author's default start downloads the 33 model files from the Hub (src/tern/artifacts.py:36-56, called from launch.py:105) - never reached, because --model points to a pre-staged local snapshot (artifacts.py:37-38 returns before any Hub import), with HF offline flags, empty credentials and a network-none container; (2) a second snapshot_download in an unused research helper (kev_vllm/clef_schema.py:501-507, no caller); (3) an exec-based rewrite of one vLLM function (kev_vllm/expert_act_quant.py:75), off in the RTX BF16 profile; (4) vLLM monkey-patches behind the author's own sha256 check of 16 vLLM source files (fail-closed); (5) subprocess calls (nvidia-smi and the server's own two child processes, list argv, no shell; src/tern/launch.py:25,33,122) inside the network-none container. Weights: matthewagi/Tern-35B-A3B at ef783c939186c1363f1de46d63564e78bb22f1c5, exactly the author's 33 inference files (69,778,082,886 bytes), fetched by us from the commit-pinned Hub URLs and sha256-verified file by file on the host, then re-hashed by the author's own verify() at every container start. Image: our own build on the pod (network on, before any sealed byte existed) from the author's digest-pinned base nvidia/cuda:12.9.1-devel-ubuntu24.04 plus the author's 196 pins re-expressed with published sha256 hashes (--require-hashes) and the pinned source tree; torch 2.13.0+cu129, vllm 0.30.0+cu129, transformers 5.19.0, triton 3.7.1, flashinfer 0.6.18.post1; image id sha256:c097beb417b5ea0d34ea5a5c144a7159d42e79dad8e9067847cce00f9e5e7ae1. FIRST MEASUREMENT (found by our own intake: Decision Index 0.4.1 entry 56, Hugging Face decision-index tag). AUTHOR-DEFAULT CONTEXT CAP: the author's server answers HTTP 422 natively above its 32,768-token limit (whole state plus all questions and options, strict >, no truncation; serve_clef_api.py:59-60,149-150); no status proxy was used. The 15 items of 76k-81k tokens (76,424-80,587) were refused exactly as pre-registered by id and position and are scored wrong, like the other self-hosted rows on this draw; the largest answered item is 1,886 tokens and none is within 64 tokens of the limit. SERIAL ONLY: the author states that answers shift with server load and batching (the Decision Index lab found 98.39% agreement with the submitted rows on 827 rows against its 98.5% bar, which it attributes to 128 concurrent requests in the submitted run versus one at a time in the check; listed under a maintainer decision); we sent strictly one request at a time, and determinism over the first 50 positions was bit-identical (50/50, 0 argmax flips). The probabilities need not equal the author's or the DI lab's (author: cross-GPU and cross-batch differences are unproven equivalent). TRAINING DATA, the author's own statement (TRAINING_DISCLOSURE.md, unverified): full-depth Qwen3.6-35B-A3B backbone with merged adapters (checkpoint 1674, about 500M tokens) and a jointly trained decision head, trained on 52 source families that include the teacher-derived jev_distill and open_jev data and train splits of public benchmarks (for example BANKING77, CLINC, HoVer, GSM8K, When2Call); 150 Decision Index-derived BBH, MMLU-Pro and When2Call requests were observed during development; the author's overlap audit lists 3,238 candidate requests with shared text among the 140,178 public Decision Index requests, 1,392 of them in index benchmarks (not proven identical decisions), and states this is not an uncontaminated held-out claim. Our own check of the exact published text hashes against our public 231 items found 0 hits; it cannot see teacher-derived knowledge or unpublished training text. The author flags 8,678 ANLI and 4,304 ContractNLI training records tagged non-commercial in intermediate dataset metadata (author: not a conclusion about the weights); the model is licensed Apache-2.0 and not marked non-commercial. SELF-REPORTED, unverified: public Decision Index score 59.01 on RTX PRO 6000, warmed serial latency median 59.48 ms on the author's card. 10 s RULE: max measured decision latency 0.092 s over all 1,500 rows; 0 decisions over 10 s.