Jev directory
CommunityJev Models & Reimplementations
JEV-9B Core AI W8A16
An independent Apple Core AI conversion of the JEV-9B text backbone with 8-bit weights and fp16 activations for on-device generation on Apple silicon; it excludes the System 1 decision head, calibration and vision.
- Reading time
- 2 min
- Published by
- Community
- Author
- silv4souz4
- Last verified
- Oct 2026
Works forJev for Document processing
Used inJev for Software & DevTools
Highlights
- 9.5 GB of Core AI assets: eight 921 MB block bundles, a 1.08 GB head and a 1.02 GB int8 embedding table, all W8A16.
- Contains only the Qwen3.5-9B text backbone: no LoRA adapter, 24-slot decision head or calibration, so it cannot make calibrated typed decisions.
- Context shrinks to 2,048 positions via a K/V ring cache, far below the original 262,144; the wrap-around is untested.
- Tested on macOS 27.2 with a 24 GB M5; the Python reference host runs about 6 tokens/s and dominates the measured speed.
- Apache-2.0 inherited from JEV-9B; converted by silv4souz4, unaffiliated with AutoTrust, TypeSafe, Alibaba/Qwen or Apple.
Quickstart
bash
hf download silv4souz4/jev-9b-coreai-w8 --local-dir jev9b
pip install numpy tokenizers
python jev9b/host/jev_coreai_host.py jev9b "The capital of France is" 24Watch out
Card reviewed 2026-10-17; quality was measured on three short texts (71 and 227 tokens) against the transformers implementation rather than on benchmarks, the ~6 tokens/s figure is an uncontrolled host-bound observation, iOS is untested, and the assets deliberately exclude the JEV System 1 decisions this catalog is about.
Questions about this entry
Short answers drawn from the linked project or page.
An independent conversion of the autotrust/JEV-9B text backbone to Apple Core AI (.aimodel) with int8 weights in blocks of 32 and fp16 activations, packaged for on-device text generation on Apple silicon.
No. The card states only the Qwen3.5-9B text backbone (embeddings, 32 layers, lm_head) is included; the System 1 LoRA adapter, 24-slot decision head and calibration are not, so it cannot produce calibrated typed decisions.
2,048 positions, implemented as a K/V ring cache, far below the original 262,144. The card notes the ring-cache wrap-around past 2,048 positions is untested because quality was only measured up to 227 tokens.
macOS 27.2 with an Apple M5 and 24 GB running the Core AI Python runtime; it was not tested on iOS. An ahead-of-time compile path for the M5 (h17g) is documented, and the iPhone 15 path is untested.
The Python reference host runs about 6 tokens/s on an M5 under uncontrolled load, but the card says the host dominates the timing and claims no latency figure for the Core AI assets themselves.
Apache-2.0, inherited from autotrust/JEV-9B and its Qwen/Qwen3.5-9B base. The conversion is by silv4souz4 and is independent of AutoTrust AI, TypeSafe AI, Alibaba/Qwen and Apple.
Fetch the repository with hf download into a local directory, install numpy and tokenizers plus the Core AI Python runtime, then run the bundled host script with a prompt and a token count; the host chunks prompts at 64 tokens.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
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 NewsMore Jev resources like this
An 8-bit MLX conversion by SirSahOl of chaoliangUNSW/Jev-Style-0.8B-Decision-v3 built for Apple Silicon native GPU inference: about 990 MB of active VRAM, a 262,144-token context and an average of 8.25 bits per weight.
Jev Models & Reimplementations#community#open-models#local
Community
A Q8_0 GGUF of knowlet's Gemma-4-12B-Unified-System-One with a native System One decision runtime and an independently fitted GGUF temperature, shipped with an F16 mmproj for the model's native image and audio projections.
Jev Models & Reimplementations#community#open-models#quantized
Community
Fluid Inference's fixed-shape Core ML conversion of the classification path of Fastino's GLiNER2.5-Decide for Apple silicon: the DeBERTa-v3-large encoder and shared label classifier score up to four decision heads per call, in fp16, 8-bit or 6-bit packages.
Jev Models & Reimplementations#community#open-models#local
Community
Serves Shanghai AI Lab's Intern-Decision-0.8B vision decision model on Apple Silicon behind a /v1/systemone-shaped endpoint: screenshots plus typed questions in, calibrated Choice, Score and Noul probabilities out, for thresholding and escalation.
Jev Models & Reimplementations#community#apple-silicon#local
An MLX float16 package of jaredpalmer/kev-4b for Apple silicon: the backbone with its adapter merged, a float32 pointer head, the tokenizer, and a package.json carrying per-file SHA-256 digests.
Jev Models & Reimplementations#community#open-models#local
Community
A calibrated image-capable 27B decision model that fits an 8 GB GPU: the JEV LoRA-and-head recipe trained on Prism ML's ternary Bonsai 2, returning typed probabilities with no generated text.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.