CommunityJev Models & Reimplementations
JevK5 GGUF
GGUF builds of the JevK5 family for llama.cpp: Q8_0 is effectively lossless (229 of 231 answers match the 4B v0.3 bf16), and every published file clears the card's 95% agreement gate.
- Published by
- Community
- Author
- alibiserikbay
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Ten published builds cover JevK5 4B in v0.2 and v0.3, JevK5-9B in v0.3 and v0.3.3, and JevK5-2B, from 2.01 GB to 9.53 GB.
- Publishing gate: a file ships only if at least 95% of its answers match bf16 on JevBench's 231 public decisions.
- 4B v0.3 Q8_0 is 4.48 GB and matches on 229/231; Q5_K_M (3.07 GB) 224; Q4_K_M (2.71 GB) 221.
- The 9B v0.3.3 Q8_0 (9.53 GB) also matches 229/231; its Q4_K_M at 222/231 with five hard items lost was not published.
- Each file has its own temperatures — for example 1.22 and 0.93 for 4B v0.3 — passed to the JevK5GGUF client.
Watch out
Only llama-server exposes the next-token log-probabilities the readout needs, so plain chat completion does not produce decisions, and each file must be paired with its own temperature pair. Q4_K_M for 4B v0.3 is noisier on the hard tier (two items lower than bf16), and the whole family remains unaffiliated with TypeSafe AI.
Questions about this entry
Short answers drawn from the linked project or page.
For a 6 GB or larger GPU, jevk5-4b-v0.3-Q8_0 at 4.48 GB: the card calls Q8_0 effectively lossless, and this file matched the bf16 model on 229 of 231 public JevBench decisions. With less memory, use the 3.07 GB Q5_K_M or the 2.71 GB Q4_K_M.
The 9B v0.3.3 Q4_K_M changed outcome on 222/231 with five hard items lost, the 9B v0.3 Q4_K_M agreed on only 218/231, and the 2B Q4_K_M dropped hard accuracy from 0.604 to 0.514. All failed the 95% agreement gate.
Each file carries its own pair: 4B v0.3 uses temperature 1.22 and knockout 0.93, 9B v0.3.3 uses 1.316 and 1.05, 9B v0.3 uses 1.049 and 1.2, 4B v0.2 uses 1.532 and 0.77, and 2B v0.2 uses 1.42 and 0.77. JevK5GGUF() with no arguments uses v0.2's values.
A current llama.cpp build with llama-server and Python 3 with NumPy for the runtime client, or the standard-library-only standalone client for up to 16 options. The pip install pins jevk5 at git+https://github.com/allebee/[email protected].
Every file was compared against its unquantized model on JevBench's 231 public decisions, scored two ways: how often it gives the same answer and its accuracy per tier. The card publishes a file only when at least 95% of its answers match.
The card says effectively: the few answers that change are borderline items and hard accuracy moves by at most two items. A v0.2 4B Q8_0 built the same way matched 230/231 when run on an M1 Pro with Metal.
The 2B Q8_0 at 2.01 GB is the card's fastest CPU option, or the 4B v0.3 Q8_0 at about 0.6 seconds per short decision.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsHacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsMore Jev resources like this
Reza2kn's llama.cpp Q4_K_M build of Jev-Omni: a 6.87 GiB multimodal decision backbone plus projector and FP32 head, measured at 8.98 GiB VRAM on an RTX 5080 Laptop GPU.
Jev Models & Reimplementations#community#open-models#local
Community
Ruiruiz30's Apple Silicon MLX conversion of Jev-Omni: 4-bit affine language weights on a Mac mini M4 16GB, about 7 GB peak Metal memory and roughly 963 ms text decisions, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
alibiserikbay's Apache-2.0 Jev alternative on Qwen3.5-4B with a merged rank-16 LoRA: one forward pass returns a probability per option, 0.784 on JevBench's public hard tier at about 13 ms per decision.
Jev Models & Reimplementations#community#open-models#local
Community
The 9B JevK5: the same v0.3 data on Qwen3.5-9B, retrained for 1.5 epochs by a pre-registered sweep, reaching index proxy 0.794 and a 0.859 hand-written hard set, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
chaoliangUNSW's v1 Jev-Style decision model in GGUF from 1.3 GB: calibrated choice, bool and score answers on Qwen3.5-2B-Base, 82.3% on five held-out task families with ECE 0.017.
Jev Models & Reimplementations#community#open-models#local
Community
The GGUF edition of TokenRhythm's NeoHorse-Jev-4B: BF16 through Q2_K quantizations that each bundle the language backbone, visual weights, decision head, and tokenizer, plus a shared runtime for text and single-image requests.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.
Jev Models & Reimplementations#open-models#training#local
Read article
Interface, size, license, claims — in that order. A selection guide across the system-one catalog, from a 706K form scorer to 27B Jev students.
Jev Practices & Patterns#open-models#evaluation#benchmarks
Read article
From the Jev community
Posts from builders shipping with Jev right now.