The story so far
Weights, every dated article, and repo commits on one rail.
8 repo commits since the weights landed
Teal is the weights (2026-09-22), grey ticks are repo commits.
JevK5
alibiserikbay Apache 2.0 out 33 days ago
4.21B parameters, text in, text out, 256K tokens of context.
Verified 2026-10-25 Rank #3932 in Decision models · score 52.1 method v1 Sources 3 Downloads 30d 120.8k · #296 of 619 tracked
Changed this week: 2 new quantized builds on the Hub (8 in all).2026-10-24
Do people use it?120.8kdownloads, 30 days#296 of 619 we track · top 48%+1 likes, downloads +867% this week24 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: 8 quantized builds and 10 Spaces.
- As a service0no API provider listed
- On own hardware8quantized builds
- Built on44 fine-tunes · 0 adapters · 0 merges
- In apps10Hugging Face Spaces
Top Spaces running it: multimodalart/jev-decision-index ♥ 802 · jmrenouard/microllm-decision-studio ♥ 1
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 16 GB card at INT4
About 2.5 GB at INT4, 5.0 GB at FP8, 10 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
What people built on it
12 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
12 derivatives on the Hub, most of them quantized
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Reference
Everything else, folded: the full tables, every source.
Benchmarks
4 rows
Benchmarks 4 rows
measured third-party leaderboard rowauthor-reported number from the model card or eval files (4)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 ECE on the teacher questions, calibrated | 0.04 | v0.3 | model author | 2026-10-07 |
| author-reported Hand-written hard set (64), accuracy | 0.77 | v0.3 | model author | 2026-10-07 |
| author-reported Held-out teacher questions (362), accuracy | 0.83 | v0.3 | model author | 2026-10-07 |
| author-reported Index proxy (16 sources, see below) | 0.73 | v0.3 | model author | 2026-10-07 |
What the card says
2 cited sentences
Summary generated, cited
It serves as an Apache-2.0 open-source alternative for typed decisions, returning option probabilities in one forward pass.[1][2] The model uses a Qwen3_5ForCausalLM architecture and can be self-hosted using the JevK5 runtime.[3][4]
- license
- “alternative to TypeSafe's Jev for typed decisions”
- architectures
- “can be self-hosted using the JevK5 runtime”
Generated 2026-10-07 15:17 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.
Use-case evidence
6 use cases
What it is good at 6 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.
- Classify and make typed decisionsstructuraldecision-model estimate p=1.00tag: decision-modeldecision catalog
- Chat and follow instructionsstructuraltag: conversationalchat template present
- Run on a laptop or phonestructuraldecision-model estimate p=0.94fits: single-gpu
- Call tools and APIsstructuralchat template: <tool_call>
- Reason through hard problemsstructuralchat template: <think>
- Read long documentsstructuralconfig: 262,144 context
Facts
20 fields
Listed in Decision models · LLMs
Facts
- JevK5 has 4.21B 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 2026-09-22, last modified 2026-09-25. (HF API, observed 2026-10-24)
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | Qwen3_5ForCausalLM |
| Model type | qwen3_5_text |
| Dtype | bfloat16 |
| Params by dtype | BF16: 4,205,751,296 |
| Size band | M |
| MoE | no |
| Context length | 262,144 (config.json max_position_embeddings) |
| Languages | en |
| License | Apache 2.0 (apache-2.0) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: supported · Qwen3_5ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Qwen3_5ForCausalLM · 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 JevK5: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 10 GB · Single GPU (<24 GB) |
|---|---|
| FP8 (1 byte/param) | 5.0 GB · Single GPU (<24 GB) |
| INT4 (0.5 bytes/param) | 2.5 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 merged; partial distillation from general LLM teachers (not Jev) |
|---|---|
| Lineage bucket | F - fine-tune of an open model (public/synthetic data) |
| Lineage evidence | model card: LoRA on Qwen3.5-4B; 17,408 of 47,460 rows labelled by teacher LLMs (Qwen3.6-27B, GPT-6 Luna), rest human-labelled public train splits |
| Base model | Qwen/Qwen3.5-4B-Base |
| DI 0.2.1 score | 38.81 (independent (DI 0.2.1 board)) |
| ECE (board) | 0.0268 |
| Median latency | 22 ms (DI board, 1x RTX PRO 6000, single process) |
| Hosted price (input) | self-hosted/M |
| Contamination notes | 980 contaminated MMLU-Pro test items (in its training data) count as wrong |
| Notes | Board raised a capacity limit ("lifted"). |
| Source | https://huggingface.co/alibiserikbay/JevK5 |
All variants
12 variants
Lineage
Base model: Qwen/Qwen3.5-4B
quantized: 8finetune: 4
Top variants (12 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| alibiserikbay/JevK5-GGUF | quantized | GGUF | — | — | apache-2.0 | alibiserikbay | 4 | 23.8k |
| crh225/plumb-4b | finetune | — | 4.21B | M | apache-2.0 | crh225 | 6 | 3.3k |
| Mr-Shmoo/sonda-1.0-4B | finetune | — | 4.21B | M | apache-2.0 | Mr-Shmoo | 0 | 995 |
| SirSahOl/JevK5-chat-mlx-16bit | finetune | mlx | 4.21B | M | apache-2.0 | SirSahOl | 0 | 324 |
| mlx-community/JevK5-chat-mlx-16bit | finetune | mlx | 4.21B | M | apache-2.0 | mlx-community | 0 | 290 |
| mlx-community/JevK5-chat-mlx-8bit | quantized | mlx | 4.21B | M | apache-2.0 | mlx-community | 0 | 239 |
| mlx-community/JevK5-chat-mlx-4bit | quantized | mlx | 4.21B | M | apache-2.0 | mlx-community | 0 | 187 |
| SirSahOl/JevK5-chat-mlx-4bit | quantized | mlx | 4.21B | M | apache-2.0 | SirSahOl | 0 | 166 |
| liodon-ai/JevK5-FP8 | quantized | -FP8 | 4.21B | M | other | liodon-ai | 1 | 75 |
| Algorythm-Canada/jevk5-0.2-mlx-4bit | quantized | mlx | 4.21B | M | apache-2.0 | Algorythm-Canada | 0 | 110 |
| SirSahOl/JevK5-chat-mlx-8bit | quantized | mlx | 4.21B | M | apache-2.0 | SirSahOl | 0 | 101 |
| Algorythm-Canada/jevk5-0.2-mlx-8bit | quantized | mlx | 4.21B | M | apache-2.0 | Algorythm-Canada | 0 | 86 |
Repository changelog
8 commits
Repository changelog 8 commits
| Date | Commit | SHA |
|---|---|---|
| 2026-09-25 | JevK5 v0.3 (4B): 17,408 teacher questions + 30,052 train-split replay items; temperature 1.22, knockout_temperature 0.93 | c4f7fdb |
| 2026-09-24 | Disclose MMLU-Pro test items in the training data (see CHANGELOG.md in github.com/allebee/jevk5) | 27d2d6b |
| 2026-09-23 | Link the GGUF builds and JevK5-2B | 05b7ad9 |
| 2026-09-23 | Clarify Jev alternative positioning and JevBench v1.4 results | b10f058 |
| 2026-09-22 | JevK5 v0.2 model card | 3c67329 |
| 2026-09-22 | JevK5 v0.2: 2x teacher data, temperature 1.532 | 844e4d0 |
| 2026-09-22 | JevK5 v0.1 | c5be5f3 |
| 2026-09-22 | initial commit | 99fe6e7 |
Similar models
3 models
Similar models
- Qwen3-4B · 4.02B same size band (M), same task (text-generation), more providers, permissive license
- Qwen3-4B-Instruct-2507 · 4.02B same size band (M), same task (text-generation), more providers, permissive license
- Qwen3-4B-Thinking-2507 · 4.02B same size band (M), same task (text-generation), more providers, permissive license
Sources
3 sources
Sources
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:43 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:17 UTC · Last verified 2026-10-25
Follow JevK5
Watch this model
- /feeds/models/alibiserikbay--JevK5.xml(RSS 2.0) · every change to JevK5
- /feeds/families/alibiserikbay--JevK5.xml(RSS 2.0) · every change across the JevK5 family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.