Fastino GLiDE 1.2 (open weights, self-hosted)v1.7 · 10 s rule

Composite (secondary)

22.0 · Composite rank #97

Cost and measurement conditions

$0.205 per 1,000 decisions (estimated)

ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-frozen 22 Oct 2026): base string Qwen/Qwen3.8-27B (the card's base_model; listed in the module's base references as qwen/qwen3.8-27b) priced as qwen/qwen3.8-27b USD 0.42/M input, USD 3.0/M output. Estimated, not charged, not a quote from the author or any host. This row does not use the operator's hosted tariff of the separate hosted rows (see disclosure). Tokens as reported by the usage block of the author's server: input_tokens = the rendered prompt of the one question (chat template and options included); output_tokens = 0 on every row (the server generates nothing: one forward pass and a 255-way linear readout). Of the 1,500 items sent, 1,485 were answered; the server-reported total of 727,293 input tokens over those 1,485 equals our pre-registered offline count (first-party transcription of the author's prompt renderer and the pinned tokenizer, made before the run). The 15 refused items (HTTP 422) carry no usage and count 0 tokens. Those 15 are the 15 standard very long items, which count for no row's cost on the common cost basis (v1.6.30). The 3 items that were sent but are not scored (see disclosure) are outside the scored set and therefore outside this row's cost basis too: the common cost basis of the scored set is 1,482 items with 724,641 input tokens, the same basis the hosted fastino-glide-api row uses. COST FIGURES (scored n = 1,497): USD 0.205364 per 1,000 decisions on the common cost basis of the scored set (1,482 items: the 1,497 scored items except the 15 very long ones that the Jev reference refused; the same basis the hosted fastino-glide-api row uses); separately, USD 0.203306 per 1,000 decisions over all 1,497 scored rows. The common-basis figure is 3.18x the official cost cap USD 0.064595 (all-rows figure 3.15x). OVER THE OFFICIAL COST CAP, DISCLOSED: the row is ranked (v1.6.16 vllm-sr-d3 precedent) and falls outside the official-caps board views by construction.

x2 + 0.15 s (assumption, not measured)

Evaluator-owned rented Verda GPU VM, one NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB (97,887 MiB), FIN-03 (Finland, EU), PAYG non-spot, instance type 1RTXPRO6000.30V, USD 2.574/h including disk. The authors' server (127.0.0.1:8000) in ONE --network none --read-only container; driver requests serial over loopback. Read-only code, model and input, empty credential variables, HF offline flags; cut proof inside the container before any sealed input (only loopback, no default route, DNS and TCP fail). Measured 23 Oct 2026 (VM reserved 08:58:49Z, scored loop started 09:05:43Z, teardown confirmed 09:10:37Z); this row measurement 277 s, scored loop 227.5 s over 1,500 requests. Frozen v1.6-regular-20261012 self-hosted input sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b, checked in the container. Raw output sha256 6ab4c6deab3cffc6e9fb7f176f52250b20d6ee936dabc9711a9858370447432a. Node-shred receipt verified; exact-id VM teardown confirms the provider instance absent and the guard closed. GPU rental is not the benchmark's estimated model tariff. Earlier attempts of this same row made no sealed contact: attempt 0 failed before any sealed byte on a runner bug (a JSON false in Python source in the import check; fixed and reviewed), and two Verda create calls answered HTTP 503 with no instance; the one exposure use (on file) was consumed by this run.

Model and serving disclosure

THE AUTHORS' OWN SERVER, run as shipped: fastino/GLiDE at a7e45896be7c96ba3354891da790c5ec99c30b6c (Hub main at the pin; the five glide/*.py files are served unmodified from the hash-verified snapshot folder, sha256 __init__.py 7e807d7d94895d436c43c2b2add0521608e208c5dde7e0016549c5b8ee5d7192, images.py 4858f30b92819abe463d845886f3240145c532fd0dfc725f5a08586e7591c616, model.py dd771ffe6c354e4a46dd1e4d8573be4faf0e8ca416010757dd2de4e5724172b2, prompt.py e47149ac980fe12c3cb24ea0ee63822d5bcde71f32d4a5cc9f37499853948338, server.py baa81296446828455b0ca1b43e58f3662b7a06ab684e5ef02d0e4fa77574030a; re-hashed inside the sealed container). Command: env PYTHONPATH=<snapshot> PYTHONDONTWRITEBYTECODE=1 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 USE_HUB_KERNELS=NO GLIDE_MODEL_NAME=glide-v1.2 GLIDE_MAX_BATCH=1 GLIDE_MAX_LEN=40000 GLIDE_API_KEY= python3 -m glide.server --model <snapshot> --host 127.0.0.1 --port 8000. bf16, PyTorch SDPA attention, one request per forward pass, temperature 1.0 from decision_config.json (sha256 df40c569d6248a8562177ea8377d47c6fbe558521f60dd14e4e8a78d04e7ce10), no kernel packages (flash-linear-attention, causal-conv1d and Hub kernels absent; checked in the container). Image: vllm/vllm-openai:v0.30.0@sha256:8a69ffad015f138d7170c4ddc429e230a3bc1c1719f67e14324749df200a4b90 used only as a CUDA 13.0 / PyTorch base (vLLM is not used); derived image sha256:f8d7c68d21d235430e49996d047c1175da52a16841fddca82fbe21fbd662f9ce, built offline from four hash-pinned wheels (transformers 5.17.0, huggingface-hub 1.33.0, fastapi 0.136.3, uvicorn 0.54.0). Python 3.12.3, torch 2.13.0+cu130, tokenizers 0.23.2, numpy 2.2.6. DEVIATIONS from the author's requirements.txt, asserted before the run: torch 2.13.0 (author 2.14.0), numpy 2.2.6 (2.5.3), fastapi 0.136.3 (0.142.2). Code gate: cleared by the benchmark owner on the disclosed patterns. DISCLOSED PATTERNS: one dormant Hub download line, snapshot_download at glide/model.py:216-218, reached only if --model is not a directory (we pass the local snapshot folder); dormant Hub-kernel loaders in the transformers Qwen3.5 code, inert with USE_HUB_KERNELS=NO and no kernels package. No trust_remote_code, subprocess, eval/exec, pickle, telemetry or request-body logging in the author code. The source and runner were reviewed before the run, findings were fixed and the final changes rechecked. Weights: Fastino Community License, public, ungated, 27 files, 52,193,616,460 bytes, every file hash-checked on the node (VERIFY OK 27 files); manifest sha256 5e9fdbcf53ffd2ae843d025e62745a145672283a0bada42bc3e90779d277d0de. The Hub main moved to e827fc56 on 23 Oct 2026 07:12Z ('Update README.md'; only README.md changed, one line making a LICENSE link absolute); the measurement stays at the pinned commit a7e45896. FIRST MEASUREMENT OF THE OPEN WEIGHTS, A NEW ROW (key fastino-glide-1.2-open; evaluator exposure and device ruling). NOT TO BE CONFUSED WITH THE TWO HOSTED ROWS: fastino-glide (hosted API, first measured 6 Oct 2026 in v1.6.1 on 600 items) and fastino-glide-api (hosted API, 22 Oct 2026 on 1,497 items; Fastino's API reports its model as glide_v2). This row is the open-weights checkpoint fastino/GLiDE at a7e45896 (version 1.2.0, engine name glide-v1.2, first uploaded 19 Oct 2026), served by us on our own rented GPU through the authors' own server; as of 22 Oct 2026 nothing published says that either hosted row served this checkpoint, and the hosted rows differ in prompt limit (65,536 vs 40,000 tokens), tariff and engine name. SCORED ON 1,497 OF THE DRAW'S 1,500 ITEMS: all 1,500 items were SENT unchanged (input hash unchanged), but 3 items Fastino had already seen through its hosted GLiNER-2.5-Decide on 8 Oct 2026 are never scored for this row (exposure rule: an item a provider has seen is never scored for that provider). The row is scored with a second invocation of the unmodified frozen scorer on a derived gold that lacks exactly those 3 ids (exclusion file sha256 31b47af58dce134a70e5cbf60aad40db0000dd398f26649070fb7c9dfd1deeb3; the derived gold's sha256 is recorded in the release proof); the other rows of this board are scored on all 1,500 items. NOT ANSWERED, counted wrong: 15 items exceed the author's 40,000-token context limit (refused above 39,992 prompt tokens; HTTP 422 'prompt exceeds the maximum context length'), exactly the 15 standard very long items (76,339-80,497 tokens) at positions 65, 169, 474, 482, 515, 573, 580, 729, 833, 899, 951, 1057, 1116, 1198, 1418, as pre-registered by id and position; never truncated. Answered: 1,485 of the 1,500 items sent with HTTP 200 (status counts 200: 1,485, 422: 15); of the 1,497 scored items, 1,482 were answered and the same 15 refused. AUTHOR'S OWN SERVER on our rented hardware, bf16, ONE REQUEST PER FORWARD PASS (GLIDE_MAX_BATCH=1). NATIVE OUTPUTS: choice argmax and probabilities, Noul P(true), score level and expectation, all from the 255-way readout softmax at temperature 1.0; no rotation or calibration file, nothing generated. OPTIONAL KERNELS OFF, FIDELITY RISK: the author's validation and all numbers in its model card used flash-linear-attention; this run used the PyTorch fallback, which the author says 'was not part of the validation', so small bf16 differences in probabilities against the author's own runs are possible and were not measured. The reason was the 10-second rule: a cold Triton cache adds 8-35 s per new shape according to the model card and the sealed container starts empty. Three package versions differ from the author's pins (see code_revision). AUTHOR EXPOSURE: the model card at the pinned revision makes no statement about training data, JevBench or decontamination; no overlap or contamination audit was run by us, and no author statement was requested or received. LICENCE NOTES: the Fastino Community License excludes evaluation, testing, benchmarking and research from 'Production' (section 6(b)); its section 4 asks for the line 'Built with Fastino GLiDE' on a related website for any Commercial Use, and the model card's introduction says 'Internal use only.' Built with Fastino GLiDE. This result does not establish that the hosted fastino-glide-api row uses the same weights. The model was under Apache-2.0 for about 50 minutes on 22 Oct 2026 before the licence changed; weights, code and configuration are byte-identical in all five revisions up to the pin (only README.md and licence files changed). 10 s RULE: 0 decisions over 10 s, maximum measured decision latency 0.485 s over the 1,500 requests, so the 10-second limit of the v1.7 share view costs this row nothing (the frozen v1.6 and v1.7-off views apply none). The longest answered prompt was 1,839 tokens (pre-registered offline count). For context, the pre-run synthetic probes (not benchmark items) took 5.3 s at 16,000 tokens and 17.06 s at the 39,992-token limit. DETERMINISM: first 50 positions run twice, 50/50 bit-identical, 0 argmax flips, maximum probability difference 0.0. CONTEXT: author-imposed limit 40,000 tokens (GLIDE_MAX_LEN, default); no truncation. The vendor's reported figures and the Decision Index figure for this checkpoint are not independently verified by us.