The story so far
Weights, every dated article, and repo commits on one rail.
30 repo commits since the weights landed
Teal is the weights (2026-09-21), grey ticks are repo commits.
metask-jev-4b-policy-mix
wayfind Apache 2.0 out 34 days ago
4.54B parameters, text in, labels out, 256K tokens of context.
Verified 2026-10-25 Rank #0 in Decision models · score 0 method v1 Sources 3 Downloads 30d 551 · #606 of 619 tracked
Do people use it?551downloads, 30 days#606 of 619 we track · below the median+0 likes, downloads −45% this week2 likesHugging Face2026-10-24
Where it's used
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
No API provider serves it, so its downloads are people running it themselves: 7 Spaces.
- As a service0no API provider listed
- On own hardware0quantized builds
- Built on00 fine-tunes · 0 adapters · 0 merges
- In apps7Hugging Face Spaces
Top Spaces running it: multimodalart/jev-decision-index ♥ 802
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 16 GB card at INT4
About 2.7 GB at INT4, 5.4 GB at FP8, 11 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
Reference
Everything else, folded: the full tables, every source.
Benchmarks
26 rows
Benchmarks 26 rows
measured third-party leaderboard rowauthor-reported number from the model card or eval files (26)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| author-reported 13 human-labeled subsets (3,880 items), macro | 78.9% percent | macro | model author | 2026-10-07 |
| author-reported Aegis · Moderate and guard content | 83.6% percent | noul | model author | 2026-10-07 |
| author-reported Commonsense QA sets · Chat and follow instructions | 87.3% percent | noul | model author | 2026-10-07 |
| author-reported ECE after per-kind temperature | 0.04 | — | model author | 2026-10-07 |
| author-reported ECE after temp | 0.04 | — | model author | 2026-10-07 |
| author-reported JevBench Score (official-methodology estimate) | 86.4 | — | model author | 2026-10-07 |
| author-reported JevBench v1.2 public 231 @4096 ctx | 80.1% percent | — | model author | 2026-10-07 |
| author-reported JevBench v1.2 — public 231 decisions, tier split | 98.6% percent | judge (original) | model author | 2026-10-07 |
| author-reported JevBench v1.2 — public 231 decisions, tier split | 100.0% percent | easy | model author | 2026-10-07 |
| author-reported JevBench v1.2 — public 231 decisions, tier split | 59.5% percent | hard | model author | 2026-10-07 |
| author-reported JevBench v1.2 — public 231 decisions, tier split | 80.1% percent | total | model author | 2026-10-07 |
| author-reported MASSIVE 14-locale dev held-out macro | 85.9% percent | macro (14 locales) | model author | 2026-10-07 |
| author-reported civil_comments | 91.3% percent | noul | model author | 2026-10-07 |
| author-reported helpsteer2 | 42.6% percent | score | model author | 2026-10-07 |
| author-reported massive-de-DE | 90.3% percent | choice | model author | 2026-10-07 |
| author-reported massive-en-US | 89.4% percent | choice | model author | 2026-10-07 |
| author-reported multinli | 90.3% percent | choice | model author | 2026-10-07 |
| author-reported p50 latency (JevBench 231, 4090) | 62.8 ms other | — | model author | 2026-10-07 |
| author-reported paws | 92.4% percent | noul | model author | 2026-10-07 |
| author-reported pubmedqa | 76.8% percent | choice | model author | 2026-10-07 |
| author-reported raw ECE | 0.11 | — | model author | 2026-10-07 |
| author-reported serving cost (owned 8×4090, dual-replica 20 QPS) | $0.0029/1k other | — | model author | 2026-10-07 |
| author-reported squad2 | 90.3% percent | noul | model author | 2026-10-07 |
| author-reported summeval-consistency | 82.6% percent | score | model author | 2026-10-07 |
| author-reported summeval-relevance | 22.9% percent | score | model author | 2026-10-07 |
Show all 26 rows (1 more)
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| author-reported vitaminc | 86.1% percent | choice | model author | 2026-10-07 |
What the card says
3 cited sentences
Summary generated, cited
Metask-Jev-4B is a 4.54B parameter calibrated typed-decision model based on the Qwen/Qwen3.5-4B architecture.[1][2] The model performs structured prediction in 16 languages using a single forward pass.[3] It is released under the Apache 2.0 license and supports a native context of 262,144 tokens.[4][5]
Generated 2026-10-07 15:19 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.
Use-case evidence
9 use cases
What it is good at 9 use cases
Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.
- Chat and follow instructionsauthor-reportedAuthor reports boolq 87.3%chat template present
- Moderate and guard contentauthor-reportedAuthor reports aegis2 83.6%
- Classify and make typed decisionsstructuraldecision-model estimate p=1.00tag: text-classificationpipeline: text-classification+1 more
- Work across languagesstructuraldecision-model estimate p=1.00tag: multilingual16 languages
- Run on a laptop or phonestructuraldecision-model estimate p=0.85name: 4bfits: single-gpu
- Understand images and videostructuraltag: image-text-to-textconfig: vision_config
- Call tools and APIsstructuralchat template: <tool_call>
- Reason through hard problemsstructuralchat template: <think>
- Read long documentsstructuralconfig: 262,144 context
Facts
21 fields
Listed in Decision models
Facts
- metask-jev-4b-policy-mix has 4.54B parameters. (safetensors metadata, observed 2026-10-24)
- Context length is 262,144 tokens (256K). (config.json max_position_embeddings, observed 2026-10-24)
- License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
- Created 2026-09-21, last modified 2026-09-22. (HF API, observed 2026-10-24)
| Pipeline tag | text-classification |
|---|---|
| Library | transformers |
| Architectures | Qwen3_5ForConditionalGeneration |
| Model type | qwen3_5 |
| Dtype | bfloat16 |
| Params by dtype | BF16: 4,539,265,536 |
| Size band | M |
| MoE | no |
| Context length | 262,144 (config.json max_position_embeddings) |
| Languages | en, zh, ja, ko, fr, es, it, ru, ar, hi, th, vi, tr, id, nl, de |
| License | Apache 2.0 (apache-2.0) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: supported · Qwen3_5ForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Qwen3_5ForConditionalGeneration · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed
vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.
Hardware fit
3 estimates
Hardware fit ESTIMATE
Weights-only memory for metask-jev-4b-policy-mix: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 11 GB · Single GPU (<24 GB) |
|---|---|
| FP8 (1 byte/param) | 5.4 GB · Single GPU (<24 GB) |
| INT4 (0.5 bytes/param) | 2.7 GB · Edge (<4 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.
Decision facts
catalog
Decision-model catalog curated
| How it was made | LoRA/adapter fine-tune |
|---|---|
| Lineage bucket | F - fine-tune of an open model (public/synthetic data) |
| Lineage evidence | board file inspection (adapter/head/weight files) + HF base_model field |
| Base model | Qwen/Qwen3.5-4B-Base |
| DI 0.2.1 score | 26.89 (independent (DI 0.2.1 board)) |
| ECE (board) | 0.1035 |
| Median latency | 57.3 ms (DI board, 1x RTX PRO 6000, single process) |
| Hosted price (input) | self-hosted/M |
| Source | https://huggingface.co/wayfind/metask-jev-4b-policy-mix |
All variants
lineage
Lineage
Base model: Qwen/Qwen3.5-4B
Repository changelog
30 commits
Repository changelog 30 commits
| Date | Commit | SHA |
|---|---|---|
| 2026-09-22 | Upload eval/figs/fig7_scatter.png with huggingface_hub | ea20fe8 |
| 2026-09-22 | Upload eval/figs/fig6_board_style.png with huggingface_hub | 74dc449 |
| 2026-09-22 | Upload eval/figs/fig5_latency.png with huggingface_hub | 494416e |
| 2026-09-22 | Upload README.md with huggingface_hub | abf13bb |
| 2026-09-22 | Upload eval/figs/fig7_scatter.png with huggingface_hub | 61ceb01 |
| 2026-09-22 | Upload eval/figs/fig6_board_style.png with huggingface_hub | 3f198d1 |
| 2026-09-22 | Upload README.md with huggingface_hub | 8036603 |
| 2026-09-22 | Upload eval/figs/fig7_scatter.png with huggingface_hub | fd3889a |
| 2026-09-22 | Upload eval/figs/fig6_board_style.png with huggingface_hub | 37e6831 |
| 2026-09-22 | Upload README.md with huggingface_hub | 531aca8 |
| 2026-09-22 | Upload eval/figs/fig8_laya_compare.png with huggingface_hub | 15ac9b0 |
| 2026-09-22 | Upload eval/figs/fig7_scatter.png with huggingface_hub | 4aad438 |
| 2026-09-22 | Upload eval/figs/fig6_board_style.png with huggingface_hub | 7439319 |
| 2026-09-22 | Upload eval/figs/fig4_calibration.png with huggingface_hub | 0a1885e |
| 2026-09-22 | Upload eval/figs/fig3_evolution.png with huggingface_hub | 376e33e |
| 2026-09-22 | Upload eval/figs/fig1_subsets.png with huggingface_hub | 693dafd |
| 2026-09-22 | Upload README.md with huggingface_hub | 9e4e1dd |
| 2026-09-22 | Upload temperature.json with huggingface_hub | 5f1f2c0 |
| 2026-09-22 | metask-jev-4b multilingual: +14 locales (MASSIVE), 16-locale checkpoint | 7694fbf |
| 2026-09-21 | selftest: English output | 81f8948 |
| 2026-09-21 | card usability: install+benchmark flow, source notes | 4e0f030 |
| 2026-09-21 | fix: judge tier filename mapping | 7e21183 |
| 2026-09-21 | add selftest.sh | c2bb9af |
| 2026-09-21 | card: selftest.sh one-liner | dfe4671 |
| 2026-09-21 | card: JevBench harness local run section (option A/B) | 5ecdd27 |
| 2026-09-21 | card: copy-paste-ready curl example | 9298d0f |
| 2026-09-21 | add serve.sh | 7c0b970 |
| 2026-09-21 | add serve.py | db16a11 |
| 2026-09-21 | card: HTTP serve section | 224d35b |
| 2026-09-21 | install.sh: transformers>=4.53 floor | 736c084 |
Papers
1 paper
Papers 1
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
Similar models
3 models
Similar models
- decider-4b · 4.21B same size band (M), same task (text-classification), permissive license
- decider-2b · 1.88B same task (text-classification), permissive license
- this-that-model-1.0 · 1.88B same task (text-classification), permissive license
Sources
3 sources
Sources
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:44 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:19 UTC · Last verified 2026-10-25
Follow metask-jev-4b-policy-mix
Watch this model
- /feeds/models/wayfind--metask-jev-4b-policy-mix.xml(RSS 2.0) · every change to metask-jev-4b-policy-mix
- /feeds/families/wayfind--metask.xml(RSS 2.0) · every change across the metask family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.