JevBench by Benchmark Heaven · released v1.4.2.2
Jev alternatives, compared on the published board
The best Jev alternative depends on your use case. Compare Jev-class decision models using the same released benchmark: Intelligence, Calibration, Speed and Cost, with each row’s evidence and openness notes.
JevBench Scores in the current top five
The Jev row is the reference; the other four rows are current alternatives. The overall score is a composite, so check the separate axes for your use case.
- 1Imajev-4B67.4I 52 · C 80 · S 91 · K 60 · ~$0.022 est.
- 2Plumb-4B65.8I 53 · C 75 · S 93 · K 56 · ~$0.030 est.
- 3decider-4b v264.1I 49 · C 75 · S 93 · K 61 · ~$0.020 est.
- 4Jev 1.13.0API63.3I 53 · C 76 · S 83 · K 52 · $0.040
- 5JevK5 v0.2.062.0I 49 · C 75 · S 91 · K 60 · ~$0.022 est.
All top-five values as a table
| Rank | System | Score | Intelligence | Calibration | Speed | Cost | Cost basis |
|---|---|---|---|---|---|---|---|
| 1 | Imajev-4B | 67.4 | 52.2 | 80.4 | 90.6 | 59.7 | estimated |
| 2 | Plumb-4B (crh225, JevK5 v0.2 + LoRA) | 65.8 | 53.0 | 75.5 | 93.5 | 55.8 | estimated |
| 3 | decider-4b v2 (Mapika) | 64.1 | 49.4 | 75.0 | 92.9 | 60.9 | estimated |
| 4 | Jev 1.13.0 (TypeSafe AI) Reference system | 63.3 | 53.1 | 76.3 | 83.3 | 52.0 | measured |
| 5 | JevK5 v0.2.0 | 62.0 | 48.9 | 74.5 | 91.1 | 59.5 | estimated |
Cost evidence is labeled by basis so estimated and announced values are not presented as measured charges.
Looking for an open source Jev alternative?
The published v1.4.2.2 aggregate contains 62 Jev-style systems with public code or weights. “Open” does not imply one shared license or unrestricted commercial use; check the source and exact component terms for your use case. Rerankers, classifiers and general LLM baselines are on the full board.
Highest-ranked open-weight entry: Imajev-4B (rank #1, JevBench Score 67.4). Is Jev itself open source?
The model column uses the published row’s description; where a parameter count was missing, the linked primary model source fills that gap. Recorded hardware is run provenance, not a minimum VRAM guarantee; missing values remain “not stated.”
Showing 62 of 62 published open-weight Jev-class rows.
| System | Model and size | Published result | License terms | Recorded run setup | Source or weights |
|---|---|---|---|---|---|
| Imajev-4B | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #1 JevBench Score 67.37 | Apache-2.0 (author adapter and server metadata) | evaluator-owned Lium GPU; exact model recorded in the run receipt | Published source or weights |
| Plumb-4B (crh225, JevK5 v0.2 + LoRA) | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #2 JevBench Score 65.84 | Apache-2.0 (weights and code; NOTICE credits JevK5, Qwen and SemIf); base Qwen3.5-4B Apache-2.0 | H100 80 GB (evaluator-owned Lium pod) | Published source or weights |
| decider-4b v2 (Mapika) | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #3 JevBench Score 64.13 | Apache-2.0 (package and weights) | RTX PRO 6000 (evaluator-owned Lium pod) | Published source or weights |
| JevK5 v0.2.0 | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #5 JevBench Score 62.04 | Apache-2.0 (code/adapter); Apache-2.0 (Qwen3.5-4B base) | re-run on a throwaway RunPod pod with the original recipe; deviations in its manifest | Published source or weights |
| Cygnet (blockbrain, frozen Gemma-4-12B-it) | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #6 JevBench Score 61.76 | shim MIT; weights Apache-2.0 with Google's Gemma Prohibited Use Policy | RTX PRO 6000 (evaluator-owned Lium pod) | Published source or weights |
| Hopper | Qwen3.5-4B plus HopitAI/hopper LoRA | Rank #7 JevBench Score 59.43 | Component-specific terms recorded in RESULT.md; submitted adapter release and Qwen base retain their respective terms | our GPU (lium.io RTX A6000 48 GB), local loopback HTTP; serial | Published source or weights |
| Winnow-12B Q8 | google/gemma-4-12B-it LoRA fine-tune, merged and exported as Q8_0 GGUF | Rank #8 JevBench Score 55.58 | Apache-2.0, including the applicable Gemma 4 base/derivative licence terms | our GPU (lium.io RTX 4090 24 GB), reached over the internet from Germany; serial, one request at a time | Published source or weights |
| reflex 4B (kshetrajna12) | Qwen/Qwen3.5-4B + kshetrajna12/reflex-qwen3.5-4b-lora | Rank #9 JevBench Score 53.99 | MIT (code, adapter); Apache-2.0 (base) | our RunPod GPU (H100 NVL 96 GB, Canada), reached over the internet from Germany | Published source or weights |
| djev (Maisa, diffusion-gemma) | inference method on google/diffusiongemma-26B-A4B-it (one structured denoising read), not a separately trained model | Rank #10 JevBench Score 52.23 | Apache-2.0 code; Google DiffusionGemma Apache-2.0 weights; no djev-specific weights | production API (api.djev.dev, free preview) | Published source or weights |
| Jev-Omni (akhilaaa3, Gemma-4-12B merged) | google/gemma-4-12B-it fine-tuned and merged, with a trained 256-way decision head | Rank #11 JevBench Score 51.34 | Apache-2.0, following Gemma 4; dataset rights stated separately by the author | our RunPod GPU (L40 48 GB, Czechia), reached over the internet from Germany | Published source or weights |
| metask-jev-4b | Qwen3.5-4B merged r16 LoRA, candidate-logit readout | Rank #12 JevBench Score 47.78 | Apache-2.0 | our GPU (lium.io RTX 5090 32 GB); serial, in-process candidate-logit inference | Published source or weights |
| SemIf, formerly OpenJev (Qwen3.5-4B, TheoLeeCJ) | Qwen/Qwen3.5-4B (frozen, BF16) | Rank #13 JevBench Score 47.69 | MIT (code); Qwen3.5 weights Apache-2.0 | RunPod RTX PRO 4500 Blackwell 32 GB (EU-RO-1) | Published source or weights |
| local-jev Qwen3.5-4B | Qwen/Qwen3.5-4B text model, zero-shot next-token letter probabilities | Rank #15 JevBench Score 46.80 | MIT code; Apache-2.0 Qwen weights | our GPU (lium.io A6000 48 GB), reached over the internet from Germany; serial, one request at a time | Published source or weights |
| system-one-open (Gemma 4 E2B LoRA on an L4) | google/gemma-4-E2B-it + LoRA Gemma 4 E2B: 2.3B effective parameters; 5.1B including embeddings. Gemma 4 model card | Rank #16 JevBench Score 45.11 | MIT (repository LICENSE; Gemma weights keep Google’s terms) | author's public demo endpoint (Modal, L4) — not a production service | Published source or weights |
| spark-s1-4b-v6 (Open Spark Jev, abhishek085) | Qwen/Qwen3.5-4B with a LoRA, read at the option-letter logits with a fitted temperature (no extra head) | Rank #17 JevBench Score 44.62 | Apache-2.0 (code and weights); base Qwen/Qwen3.5-4B Apache-2.0 | our RunPod GPU (L40 48 GB, Czechia), reached over the internet from Germany | Published source or weights |
| Malkuth-4B (newfull5, Kev post-train) | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #18 JevBench Score 44.45 | CC-BY-NC-4.0, research use only (XNLI and RACE in the training mix) | RTX PRO 6000 (evaluator-owned Lium pod) | Published source or weights |
| jqv (Qwen3-32B zero-shot) | Qwen/Qwen3-32B, bf16, read as a direct-logit classifier (no fine-tuning) | Rank #19 JevBench Score 44.35 | Apache-2.0 (Qwen3-32B weights); serving code public | our RunPod GPU (H100 NVL 96 GB, Canada), reached over the internet from Germany | Published source or weights |
| decider-35b-a3b (Mapika) | Qwen3.5-35B-A3B-Base with a trained decision readout, 34.7B parameters / 3B active | Rank #21 JevBench Score 41.18 | Apache-2.0 | our RunPod GPU (H100 NVL 96 GB), reached over the internet | Published source or weights |
| OpenSourceJev (Qwen3.5-4B Q4_K_M, native llama.cpp) | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #23 JevBench Score 40.87 | MIT (repository code); Apache-2.0 (Qwen/Qwen3.5-4B base and unsloth/Qwen3.5-4B-GGUF Q4_K_M conversion) | re-run on a throwaway RunPod pod with the original recipe; deviations in its manifest | Published source or weights |
| Malkuth-2B (newfull5, Kev post-train) | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #26 JevBench Score 38.93 | CC-BY-NC-4.0, research use only (XNLI and RACE in the training mix) | RTX PRO 6000 (evaluator-owned Lium pod) | Published source or weights |
| JEV Qwen3.5-9B Base NVFP4 | Qwen3.5-9B Base NVFP4 with compact BF16 decision head | Rank #28 JevBench Score 37.70 | Apache-2.0 entrant and upstream checkpoint | our GPU (lium.io RTX 5090 32 GB), local loopback HTTP, native FP4; serial | Published source or weights |
| OpenJev (DiffusionGemma 26B-A4B NVFP4, razorback16) | nvidia/diffusiongemma-26B-A4B-it-NVFP4 | Rank #29 JevBench Score 36.85 | Apache-2.0 (repo and weights) | RunPod RTX PRO 4500 Blackwell 32 GB (EU-RO-1) | Published source or weights |
| kev 4B (research preview) | Qwen3-4B-Base + LoRA + learned pointer head; jaredpalmer/kev-4b | Rank #30 JevBench Score 36.14 | Apache-2.0 | our RunPod GPU (GeForce RTX 3090 24 GB, community cloud CA), reached over the internet | Published source or weights |
| Decision 2B (FlyMy.AI, v59) | openbmb/MiniCPM5-2B with a trained LoRA adapter and pointer head (26.2M trainable parameters) | Rank #31 JevBench Score 35.80 | Apache-2.0 notices on the included code and the pinned base; the weights are an evaluation preview under EVALUATION-PERMISSION.md, not a cleared commercial release | our RunPod GPU (L40 48 GB, Czechia), reached over the internet from Germany | Published source or weights |
| Qwen3.5-9B Jev-like data-mix v2 | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #32 JevBench Score 35.24 | Apache-2.0 (adapter code and weights; Qwen3.5-9B base is Apache-2.0) | re-run on a throwaway RunPod pod with the original recipe; deviations in its manifest | Published source or weights |
| SimpleJev Qwen3.8-27B | Qwen3.8-27B through SimpleJev's direct-logit classifier | Rank #34 JevBench Score 34.57 | Apache-2.0 (Qwen weights); repository licence not stated | author's public demo endpoint (Featherless Classifier Demo) — not a production service | Published source or weights |
| swanOne (blockbrain, Qwen3.8-Flash-Next NVFP4) | Model details not stated in the published row. Qwen3.8-Flash family: 125B total and 6B active per token; this submission is an NVFP4 variant. QwenCloud model guide | Rank #36 JevBench Score 33.61 | patches Apache-2.0, shim MIT; weights under the Qwen licence (LICENSE-NOTICE.md) | RTX PRO 6000 (evaluator-owned Lium pod) | Published source or weights |
| open-alternative-jev (Qwen3.5-4B, IkerMoel) | Qwen/Qwen3.5-4B (frozen, BF16) | Rank #38 JevBench Score 33.24 | Apache-2.0 (code and weights) | RunPod RTX PRO 4500 Blackwell 32 GB (EU-RO-1) | Published source or weights |
| jev-local (Qwen3.5-9B) | Qwen/Qwen3.5-9B, frozen, per-option mean log-probability | Rank #39 JevBench Score 32.54 | no licence stated in the repository (public code); Apache-2.0 base weights | our RunPod GPU (H100 NVL 96 GB, Canada), reached over the internet from Germany | Published source or weights |
| Decision Fast (FlyMy.AI, v53a) | Qwen/Qwen3-0.6B-Base with a trained LoRA adapter and pointer head (10.6M trainable parameters) | Rank #40 JevBench Score 32.49 | Apache-2.0 notices on the included code and the pinned base; the weights are an evaluation preview under EVALUATION-PERMISSION.md, not a cleared commercial release | our RunPod GPU (L40 48 GB, Czechia), reached over the internet from Germany | Published source or weights |
| decider-2b (Mapika) | Qwen3.5-2B-Base with a trained decision readout, 1.9B | Rank #41 JevBench Score 30.74 | Apache-2.0 | our RunPod GPU (H100 NVL 96 GB, Canada), reached over the internet from Germany | Published source or weights |
| jeff (Logan Markewich, GLiFormer 400M) | GLiFormer large (knowledgator/gliformer-large-v1, ~400M) behind a TypeSafe-compatible /v1/systemone server | Rank #42 JevBench Score 30.58 | MIT (code); GLiFormer weights per their model card | AMD Ryzen 5 3600 (Sandy), 4 threads | Published source or weights |
| Laya (Convai Innovations, ModernBERT-large 421M) | ModernBERT-large encoder + option-marker decision head, 421M, RLCD-trained | Rank #43 JevBench Score 30.25 | Apache-2.0 | AMD Ryzen 5 3600 (Sandy), 4 threads | Published source or weights |
| Standard One 8B (Standard Thinking) | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #45 JevBench Score 29.07 | Apache-2.0 (jev-adapter server, LoRA and merged weights; base Ministral 3 8B Apache-2.0) | RTX PRO 6000 (evaluator-owned Lium pod) | Published source or weights |
| lev-350m (Franck Verrot, LFM2.5-350M) | LiquidAI/LFM2.5-350M with a LoRA and a 6.5M-parameter pointer head (a kev clone) | Rank #46 JevBench Score 28.50 | Apache-2.0 (code); weights under LiquidAI's LFM1.0 licence, following the LFM2.5-350M base | our RunPod GPU (L40 48 GB, Czechia), reached over the internet from Germany | Published source or weights |
| openjev-sglang (Qwen3.6-35B-A3B on SGLang) | Qwen3.6-35B-A3B | Rank #47 JevBench Score 27.65 | no licence file in the repository as of 2026-09-19; Qwen3.6 weights keep their own terms | author's public demo endpoint (Modal) — not a production service | Published source or weights |
| Von (wfzyx, Option-Marker 395M) | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #48 JevBench Score 27.48 | Apache-2.0 code; Apache-2.0 weights (answerdotai/ModernBERT-large base) | re-run on a throwaway RunPod pod with the original recipe; deviations in its manifest | Published source or weights |
| kev 8B (research preview) | Qwen3-8B-Base + LoRA + learned pointer head; jaredpalmer/kev-8b | Rank #51 JevBench Score 25.56 | Apache-2.0 | our RunPod GPU (GeForce RTX 3090 24 GB, community cloud CA), reached over the internet | Published source or weights |
| JevOne | Qwen3.6-35B-A3B BF16 with JevOne bidirectional option-logit mapping | Rank #52 JevBench Score 25.51 | Component-specific JevOne/Qwen/SGLang terms recorded in RESULT.md | our GPU (lium.io RTX PRO 6000 Blackwell 96 GB), local loopback HTTP, TP1; serial; independently validated one-card condition | Published source or weights |
| typecastlm (Mikhail Gribov, Qwen3.5-4B computed head) | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #53 JevBench Score 25.28 | Apache-2.0 (package and weights) | RTX PRO 6000 (evaluator-owned Lium pod) | Published source or weights |
| SimpleJev Qwen3.6-35B-A3B | Qwen3.6-35B-A3B through SimpleJev's direct-logit classifier | Rank #54 JevBench Score 24.86 | Apache-2.0 (Qwen weights); repository licence not stated | author's public demo endpoint (Featherless Classifier Demo) — not a production service | Published source or weights |
| kev 0.6B (research preview) | Qwen3-0.6B-Base + LoRA + learned pointer head; jaredpalmer/kev-0.6b | Rank #55 JevBench Score 24.75 | Apache-2.0 | our RunPod GPU (GeForce RTX 3090 24 GB, community cloud CA), reached over the internet | Published source or weights |
| system-one (Qwen3-8B, Sean Goedecke) | Qwen/Qwen3-8B (frozen, BF16), the model of the author's demos | Rank #57 JevBench Score 23.37 | no licence file in the repository as of 19 Sep; Qwen3 weights Apache-2.0 | RunPod RTX PRO 4500 Blackwell 32 GB (EU-RO-1) | Published source or weights |
| LitJev (Qwen3.8-27B) | Qwen/Qwen3.8-27B, frozen, read at the output head | Rank #59 JevBench Score 19.51 | Apache-2.0 (code); Apache-2.0 base weights | our RunPod GPU (H100 NVL 96 GB, Canada), reached over the internet from Germany | Published source or weights |
| openJev Verdict 1.4 | GLiClass ModernBERT-base fine-tuned decision model, 151M; v1.4 fixed inference engine | Rank #60 JevBench Score 19.00 | Apache-2.0 | AMD Ryzen 5 3600 (Sandy), 4 threads | Published source or weights |
| kev 0.5B | Qwen2.5-0.5B + LoRA + learned pointer head; jaredpalmer/kev-0.5b | Rank #61 JevBench Score 18.88 | Apache-2.0 | our RunPod GPU (GeForce RTX 3090 24 GB, community cloud CA), reached over the internet | Published source or weights |
| Bespoke Nimble 9B (Bespoke Labs) | bespokelabs/Bespoke-Nimble-9B (LoRA, adapter unchanged since 93ec5d6), merged into Qwen/Qwen3.5-9B@c202236 with the author's PEFT safe-merge | Rank #62 JevBench Score 18.66 | Apache-2.0 (weights); repository without a licence file as of 19 Sep | our RunPod GPU (A40 48 GB, Canada), reached over the internet from Germany | Published source or weights |
| openJev Verdict (heman10x, ModernBERT-base 151M) | GLiClass ModernBERT-base (knowledgator/gliclass-modern-base-v2.0) fine-tuned, 151M | Rank #64 JevBench Score 18.09 | Apache-2.0 | AMD Ryzen 5 3600 (Sandy), 4 threads | Published source or weights |
| reflex-27b (Qwen3.8-27B) | Qwen3.8-27B, frozen, direct-logit readout averaged across two option orders | Rank #66 JevBench Score 17.84 | MIT code; Apache-2.0 Qwen weights | our RunPod GPU (H100 NVL 96 GB), reached over the internet | Published source or weights |
| djev (thinking) | google/diffusiongemma-26b-a4b-it, BF16; full generation with thinking enabled | Rank #68 JevBench Score 15.20 | Apache-2.0 | our GPU (lium.io H200 141 GB), reached over the internet from Germany; serial, one request at a time | Published source or weights |
| OpenJev (thinking, BF16) | google/diffusiongemma-26b-a4b-it, BF16; OpenJev think=512 | Rank #70 JevBench Score 14.83 | Apache-2.0 | our GPU (lium.io H200 141 GB), reached over the internet from Germany; serial, one request at a time | Published source or weights |
| Qwen3.5-0.8B Decision Model (Mourad Ghafiri) | Qwen3.5-0.8B-Base fine-tuned as a JevLite decision model with per-question calibration | Rank #71 JevBench Score 14.54 | MIT code and training data; Apache-2.0 model weights | Ryzen 5 3600, 4 threads | Published source or weights |
| open-jev-deberta-v3-large (local CPU) | microsoft/deberta-v3-large DeBERTa-v3-large: 304M backbone plus 131M embedding parameters. Microsoft model card | Rank #73 JevBench Score 12.65 | Apache-2.0 (model card); DeBERTa-v3 keeps its own terms | our CPU (2 threads, Ryzen 5 3600) | Published source or weights |
| smalljev semantic-v9 | MiniCPM5-2B-Base, 2.5B dense, with LoRA and native semantic decision heads | Rank #74 JevBench Score 12.31 | Apache-2.0 | our GPU (lium.io A6000 48 GB), reached over the internet from Germany; serial, one request at a time | Published source or weights |
| Open-Jev 9B (Zefan Cai) | Qwen3.5-9B plus rank-8 LoRA and trained scalar decision head | Rank #77 JevBench Score 11.20 | MIT (loader); Apache-2.0 (adapter and pinned Qwen base); CC0-1.0 public training projection | our RunPod GPU (H100 80GB HBM3), reached over the internet | Published source or weights |
| Open-Jev 2B (Zefan Cai) | Qwen3.5-2B plus rank-8 LoRA and trained scalar decision head | Rank #78 JevBench Score 9.98 | MIT (loader); Apache-2.0 (adapter and pinned Qwen base); CC0-1.0 public training projection | our RunPod GPU (H100 80GB HBM3), reached over the internet | Published source or weights |
| CLM-8B (Contrastive-LM, clm-latest) | Model details not stated in the published row. Parameter size not stated in the published row. | Rank #80 JevBench Score 8.57 | Apache-2.0 (code and CLM-v0.1-8B head); Qwen3-8B encoder Apache-2.0 | RTX PRO 6000 (evaluator-owned Lium pod) | Published source or weights |
| SimpleJev (Qwen3.5-0.8B, CPU) | Qwen3.5-0.8B through SimpleJev native assistant-prefill option-logit scorer | Rank #81 JevBench Score 7.46 | Apache-2.0 server; Apache-2.0 Qwen3.5-0.8B checkpoint | Sandy shared CPU (4 inference threads) | Published source or weights |
| verdict-small (Manavarya09, multilingual-e5-small 118M) | intfloat/multilingual-e5-small (118M multilingual bi-encoder) fine-tuned on a typed-decision mix; each option is scored against the rendered state by cosine similarity, with temperature scaling and a conformal abstain set on top. One encoder pass per option, no tokens generated. | Rank #84 JevBench Score 5.69 | Apache-2.0 (verdictml code and the Manav2op/verdict-small checkpoint over intfloat/multilingual-e5-small) | AMD Ryzen 5 3600 (Sandy), 4 threads | Published source or weights |
| Mirror | Mirror DeBERTa-v3-large 436M classification-span scorer | Rank #86 JevBench Score 2.11 | Apache-2.0 wrapper and Mirror release; upstream DeBERTa/model assets retain their own terms | hosted submission endpoint, serial from Germany; strict 512-token context rejection | DeBERTa-v3-large base model card The submitted Mirror head/weights were unavailable at the published repository URL when checked. This link is only the DeBERTa base model, not the complete measured Mirror system. |
| Certo v1 (AltSlate Labs) | ModernBERT-large with a per-option query/scoring head, ~400M parameters | Rank #90 JevBench Score 0.00 | MIT | our RunPod GPU (GeForce RTX 3090 24 GB, community cloud), reached over the internet from Germany | Published source or weights |
| Open Jev JSON Canvas (JoshuaSP) | google/diffusiongemma-26B-A4B-it BF16, one-step JSON canvas | Rank #91 JevBench Score 0.00 | MIT code; Apache-2.0 DiffusionGemma weights | our lium.io H100 80 GB; local in-process; serial; seed 0; one denoising step | Published source or weights |
Mirror remains listed for completeness. Its published repository returned 404 when checked, so the table links only to Microsoft’s DeBERTa-v3-large base model card.
See the use-case chooser, including accuracy, speed, cost and self-hosting evidence. · Jev vs Laya, a CPU-measured open model
Frequently asked questions
- What does JevBench compare?
- JevBench compares published Jev-class decision systems across Intelligence, Calibration, Speed and Cost. This page uses the released v1.4.2.2 aggregate.
- Which open-weight Jev alternative scores highest?
- Imajev-4B is the highest-ranked open-weight entry in v1.4.2.2: rank 1 overall with a JevBench Score of 67.4. Imajev-4B leads the JevBench Score at 67.37, ahead of Plumb-4B (65.84). The v1.4.2 scoring code and earlier measurement rows are unchanged. The score is a composite, not an accuracy percentage or a guarantee for your workload.
- What counts as an open-weight Jev alternative on this page?
- All 62 Jev-style rows in v1.4.2.2 whose code or weights are public, with a license note and source link in the published row. The license text is kept per row because code, adapters, base weights and datasets can have different terms, including non-commercial ones.
- Does the recorded hardware show a minimum deployment requirement?
- No. The table reports the setup used for each benchmark run. 8 of these rows were measured on four CPU threads; most others ran on a rented GPU or the author’s own endpoint. It is not a minimum VRAM or hardware guarantee for another revision, quantization, context length or serving stack.
- Are cost and speed values directly measured?
- The board labels cost evidence as measured, estimated or announced. Self-hosted rows usually carry an estimated cost from a comparable hosted price, and rows served on our own hardware carry a published speed adjustment (such as ×2 + 0.15 s) that is an assumption, not a measurement.