Jev-Style

[图片]

Jev-Style

Small decision models that run on your own machine. Text and images in, a calibrated probability for every option out.

Website · Models · Live demo · PyPI · Agent skills · MCP · Changelog · 简体中文

[PyPI] [CI] [Weights: 0.8B · 2B · torch · MLX · GGUF] [Demo: 2B v4 on HF Spaces] [MCP Registry] [License: Apache-2.0]

Created and maintained by Chaoliang Yan. An independent project, not affiliated with TypeSafe or Jev (details).

Jev-Style is a family of small decision models (0.8B and 2B, fine-tuned from Qwen3.5) and the toolkit that puts them to work. Give a model text or JSON, optional images and typed questions. It reads the input once and returns a calibrated probability for every option, in one forward pass, on your laptop. Use it to route support tickets to the right team, pick the tool or model an agent should call, tag topics, flag phishing, check receipts and screenshots, or let a coding agent's tool calls through only when a local model agrees. No GPU, no API key and no training needed.

[The Playground answering a support-ticket request, then the agent-approval demo allowing, asking about and denying tool calls]

Quick start

Try it first with nothing installed: the 2B Space runs Jev-Style-2B-Decision-v4 (text and images) and the 0.8B Space runs the 0.8B.

Locally, you need Python 3.10 or newer. In a virtual environment, install one backend for your machine:

pip install -U "jev-style[mlx]"      # Apple silicon (text)pip install -U "jev-style[torch]"    # NVIDIA GPU, CPU or image input; Intel Macs, Linux and Windows (also works on Apple MPS)

The bare pip install jev-style installs only the HTTP client, the fake engine and the MCP server; running a model needs one of the extras above. [mlx] pins mlx-lm to 0.31.3 on purpose (the 2B MLX runtimes check the mlx-lm source they patch and refuse any other version), so give jev-style its own virtual environment, or use uv tool install, if other MLX tools in the same environment need a newer mlx-lm.

Start the server and open the local Playground. The first start downloads the 0.8B weights (about 1.5 GB); later starts reuse them.

jev-style serve                      # add --release 2b for the 2B v4 (images need --backend torch)

In a second terminal, make one decision:

jev-style decide "I was charged twice" --url http://127.0.0.1:8765 --choice "Which team?::billing,shipping,returns"

Or from Python, in one line (loads the model in-process, or uses the server in JEV_STYLE_URL):

import jev_style
jev_style.classify("Where is my parcel? It was due Monday.", ["billing", "shipping", "tech"])# {'label': 'shipping', 'confidence': ..., 'probabilities': {'billing': ..., 'shipping': ..., 'tech': ...}}

The server speaks the public POST /v1/systemone request shape, so clients written for Jev-compatible servers, including the official SDKs, can point at http://127.0.0.1:8765 with any API key string. For the command line tool on its own: uv tool install "jev-style[all]" on Apple silicon, uv tool install "jev-style[torch]" elsewhere.

Why Jev-Style

• Three question types in one request. Yes/no (noul), multiple choice (choice, up to 255 options) and ordered ratings (score, 2 to 10 levels). The model reads the text once and answers every question about it.
• Probabilities, not just labels. Each release ships temperatures fitted on held-out data, so your code can act on confident answers and send the rest to a person or an LLM.
• Images (2B v4). Photos, screenshots, receipts, charts and scanned forms in the state, sent as assets the way TypeSafe's API takes them. The PyTorch build reads them; the MLX and GGUF builds read text only.
• Long inputs. Up to 32,768 tokens per call with the 2B v4 (25,600 with v3). Nothing is truncated: an input that is too long is rejected with an error that says so.
• Runs locally, small. 0.53 GB for the 0.8B at 4-bit; about 0.15 to 0.2 s for a short request with MLX on an M1 Max, after the first call. Nothing leaves your machine.
• 51 languages evaluated. Training covers 19 languages. On MASSIVE intent the 0.8B beats Laya's multilingual checkpoint in all 51 evaluated languages.
• Built for agents. Six agent skills, an MCP server in the official registry, a Claude Code tool-call guard and confidence cascades, all working with Claude Code, Codex, Cursor and other agents.
• Measured, with the protocol published. jev-style eval scores your own labelled data; every benchmark number here links to a model card with its protocol and confidence intervals.

Models

Start with the default 0.8B to try things: 1.5 GB, fast on a laptop. Move to the 2B v4 for accuracy, images and 32,768-token inputs. --release 2b loads the v4; --release 2b-v3 keeps the previous 2B.

Decision Index. Since 2026-11-04 the 2B v4 is listed on the public Decision Index 0.3 board as jev-style-2b-v4, replacing the 2B v3 entry (34.05): Public 44.37, Full 34.53, Breadth 41.75, and on the Vision board Public 57.55 / Full 51.94. The 0.8B v3 is on the same board at Public 17.0 / Full 12.5. The Full score also counts the maintainers' private test sets, which is why it is lower than the Public score.

How to read the JevBench numbers. The 2B v4 JevBench score is one self-run with the official harness through this package's server (float32, Apple MPS; 95 % CI 68.9 to 80.0 %), not a JevBench board entry. Several benchmarks' official train splits are in the 2B v4's training data, so those results are not zero-shot. The 2B v3 numbers are single pre-declared runs with its GGUF F16 build: on JevBench it is the highest among the Qwen3.5-2B-family systems on the v1.4.1 board (decider-2b 71.0 %, open-jev-zefan-2b 64.5 %), but the lead over decider-2b is inside the 95 % confidence interval, 42 of the 82 board systems score higher, and hosted Jev is well ahead. Each model card gives every protocol: 2B v4, 2B v3, 0.8B v3.

Where it fits

The other projects' cells come from their READMEs on 2026-11-03; corrections are welcome. Use hosted Jev when you want the highest accuracy and your data may leave the machine; Kev when you want to train your own sizes up to 27B; Laya when you need the fastest CPU answers in many languages; Jev-Style when you want a pip install, images, long inputs and the agent tooling below.

Use it from Python

For several questions about one text, and to choose the engine yourself:

from jev_style import JevStyle, choice, noul
js = JevStyle.from_pretrained("chaoliangUNSW/Jev-Style-2B-Decision-v4-MLX")      # the 2B v4 in-process; the repo picks the backend# js = JevStyle.from_pretrained("chaoliangUNSW/Jev-Style-0.8B-Decision-v3-MLX")  # the 0.8B# js = JevStyle(release="2b")                                                    # the 2B v4, backend chosen for this machine# js = JevStyle()                                                                # the default release (0.8B)# js = JevStyle(base_url="http://127.0.0.1:8765")                                # or a running serverout = js.decide("I was charged twice. Please fix this ASAP.", {    "billing": noul("This ticket is about billing."),    "team":    choice("Which team should handle this?", {"billing": "charges, refunds, invoices",                                                          "shipping": "delivery, lost parcels", "tech": "bugs, outages"}),    "refund":  noul("The customer asks for money back."),})print(out["answers"]["billing"]["noul"], out["answers"]["team"]["choice"])

The same call covers other routing decisions, for example flagging phishing and tagging the topic of an email, or picking the tool an agent should call:

email = js.decide(email_text, {    "phishing": noul("The message tries to trick the reader into revealing credentials or paying."),    "topic":    choice("What is the message mainly about?", ["invoice", "delivery", "account security", "other"]),})tool = js.decide(user_request, {    "tool": choice("Which tool should the agent call?", {"web_search": "fresh facts from the web",                                                         "calendar": "dates and meetings",                                                         "python": "arithmetic, data, files"}),})

from_pretrained takes the main repo (PyTorch), -MLX (precision="8bit" for the 2.00 GB 2B or 0.80 GB 0.8B weights) or -GGUF (quant="Q4_K_M", needs the release's scorer, see Backends). The client is not tied to this model: JevStyle(base_url=...) works with any server that implements POST /v1/systemone, including the hosted API. JEV_STYLE_API_KEY (the key of your local jev-style serve --api-key-env, or of the hosted API) is sent only to loopback servers and to the server named in JEV_STYLE_URL; for any other server pass api_key= (--key NAME=ENV for jev-style eval, --api-key-env NAME for jev-style decide --url, api_key_env for a cascade tier), so a key is never sent to a third-party server by accident.

Images go in the state by reference, with the image itself under assets (2B v4, PyTorch build):

from jev_style import JevStyle, image, noul
js = JevStyle.from_pretrained("chaoliangUNSW/Jev-Style-2B-Decision-v4")         # PyTorch: text and imagesout = js.decide({"receipt": "r1", "policy": "Meals over 50 EUR need a manager approval."},                {"approval": noul("Does this receipt need a manager approval?")},                assets={"r1": image("receipt.png")})                            # PNG or JPEG, 280 to 1,568 px per sideprint(out["answers"]["approval"]["noul"])

From the shell, with or without a server:

jev-style decide "Refund still missing after 3 weeks" --url http://127.0.0.1:8765 \  --choice "Which team?::billing,shipping,returns" --noul "The customer is angry"# without --url (or JEV_STYLE_URL) it loads its own copy of the model (add --release 2b for the 2B v4)

--api-key-env NAME sends the key in that environment variable to a --url that is not loopback; --verify checks the model files against their manifest before an in-process load.

The API, by example

With the server running, open http://127.0.0.1:8765 for the Playground and the demos (agent action approval, Snake, Chinese and 51 languages), or send a ticket:

curl -s localhost:8765/v1/systemone -H 'content-type: application/json' -d '{  "state": "Shoes arrived two weeks late and in the wrong size. Also I see two charges on my card.",  "questions": {    "department":  {"type": "choice", "instructions": "Which team should handle this?",                    "criteria": {"returns":  "Exchanges, refunds, wrong or damaged items",                                 "shipping": "Delivery status, delays, lost packages",                                 "billing":  "Charges, invoices, payment problems"}},    "escalate":    {"type": "noul",  "instructions": "Does this need urgent human attention?"}  }}'

These two answers are from the actual response of the 0.8B (the default release) with MLX on an Apple M1 Max, probabilities rounded (that run also asked a third question, so its usage and timing fields are left out):

{  "model": "jev-style-0.8b-decision-v3",  "answers": {    "department":  { "type": "choice", "choice": "billing", "confidence": 0.46,                     "probabilities": { "returns": 0.07, "shipping": 0.29, "billing": 0.64 } },    "escalate":    { "type": "noul", "noul": 0.45 }  },  "backend": "mlx"}

The ticket raises both a billing problem and a late delivery, and the probabilities show it: billing 0.64, shipping 0.29. That is why the model returns probabilities rather than a single label. Your code can act on the confident answers and hand the rest to a person. The full request and error reference is under API.

Agent skills

The skills live in skills/. Each one tells a coding agent how to do one job from start to finish, and ends with a check that the job worked.

npx skills add lawrence3699/jev-style                    # pick skills interactivelynpx skills add lawrence3699/jev-style --skill '*' -y     # all six

In Claude Code you can also install everything as plugins:

/plugin marketplace add lawrence3699/jev-style/plugin install jev-style@jev-style          # the six skills + MCP tools/plugin install jev-style-guard@jev-style    # the PreToolUse guard

Claude Code guard

Before Claude Code runs a Bash, Write, Edit, Read, WebFetch or MCP call, jev-style guard sends the call to the local model. It asks whether the call is destructive, exfiltrates data, touches secrets or goes outside the project, plus a 0 to 4 risk score, and turns the answers into allow / ask / deny. Regex hard rules in code can only make a verdict stricter. If the server is down, times out, rejects the request or answers without a usable value for every question, the verdict is ask, never a silent allow.

jev-style guard --check "git push --force origin main"# decision: DENY   (source: rule, 443 ms)  destructive 0.79 ...jev-style guard-replay --url http://127.0.0.1:8765     # the 49 bundled labelled tool calls

On the 49 bundled tool calls, written and labelled by hand, the default config with the 0.8B v3 (the default release; the 2B releases were not measured on this set) agrees with the labels on 77.6 %. No call labelled deny was allowed. 2 of the 16 calls labelled ask were allowed. The model alone, with no hard rules, agrees on 61.2 %. Use it as a second line of defence, not as a sandbox. The skill covers installing, dry runs and tuning.

MCP server

claude mcp add jev-style --scope user -- jev-style mcp       # Claude Code

jev-style mcp is a thin stdio server that forwards to the running jev-style serve, so one copy of the model serves every client. It is listed in the official MCP Registry as io.github.lawrence3699/jev-style (runs as uvx jev-style mcp; set JEV_STYLE_URL if your server is not on http://127.0.0.1:8765). It provides decide (several questions about one input), noul, choice, score and model_info. Its instructions to the client give the served model's context budget and say whether the state stays on this machine: with --allow-remote --url https://... every state goes to that server, and the instructions say so. Setup for Codex, Cursor and Claude Desktop is in the skill.

Evaluate on your own data

The best evidence is your own labelled examples. jev-style eval takes JSONL in the request format, with a label on each question:

jev-style eval examples/support_tickets.jsonl --url http://127.0.0.1:8765

question                   n     acc   brier    ece   automate @1% / 5% / 10% errorteam                      30   93.3%   0.077  0.123    90.0% (p>=0.65) /  93.3% (p>=0.57) / 100.0% (p>=0.44)escalate                  30   96.7%   0.107  0.125    80.0% (p>=0.71) / 100.0% (p>=0.54) / 100.0% (p>=0.54)mood                      30   73.3%   0.433  0.174    40.0% (p>=0.56) /  40.0% (p>=0.56) /  46.7% (p>=0.54)

The 30 example tickets were written by hand for this repository; they demonstrate the format and are not a benchmark. Read automate @5%: 93.3% (p>=0.57) like this: if you act only when the top probability is at least 0.57, the model handles 93.3 % of tickets with at most 5 % errors among them, and everything else goes to a person or an LLM. The report also refits a temperature on half the rows and shows its effect on the other half. See the skill for building a proper evaluation set.

jev-style eval examples/support_tickets.jsonl --server bf16=local:mlx --server q8=http://127.0.0.1:8799

90 questions answered by every engine; the table and the differences use only those.
engine             answered failed rows     acc   brier log loss    ece   automate @5% errorbf16                     90           0   87.8%   0.206    0.369  0.125    80.0% (p>=0.54)q8                       90           0   87.8%   0.206    0.370  0.117    78.9% (p>=0.54)
difference to bf16 (95 % paired bootstrap over rows; brier: lower is better):  q8               accuracy  +0.0 pts [+0.0, +0.0]   brier +0.001 [-0.001, +0.002]

Recipes

Clone this repository to run the three complete examples. Each prints JSON and makes no changes to tickets or payments.

The text examples include 12 hand-written synthetic tickets for checking the workflow; thresholds are illustrative. Evaluate on your own held-out data before automating decisions. The Semantic Router example connects DecisionRouter (semantic-router==1.1.1) to the running server without an external API key: python examples/semantic_router_local.py "Where is my parcel?". Guides with more context: local classification and routing, support ticket classification, choosing between a classifier, embeddings, an LLM and a decision model, comparing the models.

Cascades

A cascade answers each question with the smallest model that is confident enough and sends only the rest to a larger one. Tier 1 answers every question in one call; a question whose confidence (k · p_max − 1) / (k − 1) (for noul |2 · P(true) − 1|) is below the tier's threshold, or that the tier failed on, goes to the next tier in one call with the same state. The top tier's answer is final; if it fails, the last answer that succeeded is used, and only if every tier failed is the error returned. Answers keep the usual schema and add tier and tier_model; timing.tiers reports each tier's calls.

--cascade (for serve, decide and download), JevStyle(cascade=...) and the eval engine NAME=cascade:... take either the name of a cascade that ships with the package or a cascade file. jev-style releases lists the named cascades with their thresholds, tiers and pinned revisions.

cascade-9b (model id jev-style-cascade-9b) runs Jev-Style-2B-Decision-v3 in-process (PyTorch, float32, CUDA graphs) and sends a question whose confidence is below τ = 0.75 to the third-party JevK5-9B v0.3.3 by alibiserikbay (Apache-2.0), run in the same process through its own runtime, allebee/jevk5, at a pinned revision. τ was frozen on a 4,992-question calibration set (sha256 580962e4b2bc86b32b42731eda42536f57321c7d1ca9ac1e3b970bf7dbd7e2d2) with a rule written down before any result: the lowest expected latency whose calibration accuracy stays within 1.0 point of JevK5-9B alone. Details and every number: chaoliangUNSW/Jev-Style-Cascade-9B.

We do not claim that cascade-9b is more accurate than JevK5-9B on its own; what the cascade changes is how many questions reach the 9B at all. It needs an NVIDIA GPU with room for both models (on an RTX 5090 with 32 GB: 29.7 GB after start-up, at most 31.8 GB while serving, with PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True), PyTorch and the jevk5 package, which is not on PyPI. JevK5-9B's model card says that part of its training labels came from OpenAI's GPT-6 Luna under OpenAI's terms; check that those terms suit your use.

pip install "jev-style[torch]" "jevk5[fast] @ git+https://github.com/allebee/[email protected]"jev-style download --cascade cascade-9b      # both tiers; the JevK5-9B snapshot is checked against its SHA256SUMSjev-style serve --cascade cascade-9b --backend torch --device cuda

jevk5-serve --model alibiserikbay/JevK5-9B --port 8090        # on the GPU box (see allebee/jevk5)jev-style serve --cascade examples/cascade/glue-2.json --backend torch --device cuda

Backends and speed

--release works the same for serve, decide, download and mcp --model; JEV_STYLE_RELEASE=2b sets it for every command and for the Python client.

CUDA graphs. On an NVIDIA GPU the PyTorch backend records CUDA graphs when it starts and then replays them, instead of launching every kernel of a call one by one. The 2B v4 on an A100 80GB in bfloat16, 750 Decision Index 0.3 public requests one at a time: median 16.4 ms with CUDA graphs, 71.5 ms without (its model card has the protocol and the parity checks). The v3 releases on an RTX 5090 in float32, 4,992 calibration questions against the same runtime without graphs:

Under --cuda-graphs auto the v3 releases use graphs whenever they run on CUDA (recording adds about 7 to 11 s to start-up); the 2B v4 runtime decides itself: on with --dtype bfloat16, off in float32, where its forward is compute-bound and recording costs about a minute for little gain. --cuda-graphs on|off, JEV_STYLE_CUDA_GRAPHS=1|0 or JevStyle(..., cuda_graphs=True|False) overrides that. Inputs longer than 4,096 tokens, and several questions over a long shared state, take the ordinary path. It only applies to the PyTorch backend on CUDA.

API

POST /v1/systemone with {"state": text | object | array, "questions": {id: question}}, plus "assets" for images (2B v4):

Choice confidence = (k · p_max − 1) / (k − 1) for k options: 0 when the probabilities are uniform, 1 when one option takes all of them. Score confidence is TypeSafe's ordinal formula, which also counts how far probability sits from the most likely level m: max(0, 1 − Σ pᵢ·|i − m| / MAD_unif), with MAD_unif the same average distance for an even spread. Noul answers carry no confidence. Images follow TypeSafe's API: a string value in state names an entry of assets ({"type": "image", "format": "png" | "jpg", "data_base64": ...}, or a multipart file part), the request carries TypeSafe-Danger-Zone-Images: true, and each image is a still PNG or JPEG of 280 to 1,568 px per side; a model that reads text only answers 400 images_not_supported. Bodies may be compressed with Content-Encoding: gzip, x-gzip or zstd (the official SDKs compress by default), and the request rules match TypeSafe's API. The other routes are GET /v1/models, GET /healthz and the Playground at /. Errors look like {"error": {"code", "message", "question"?}} and use HTTP 400 (max_tokens_exceeded, request rules), 413, 415, 422 (invalid_json, invalid_request, invalid_question), 401 (unauthorized), 403 (forbidden_host, forbidden_origin), 404 or 500. The hosted api.jevstyle.com answers with a flat body instead, {"error": "<code>", "message": "..."}, and reports over-long input as 413 input_too_large; the Python client, the guard and the MCP server read both shapes, and JevStyleError.input_too_long is true for either code. Start the server with --api-key-env NAME to require Authorization: Bearer <key>. The local server answers only requests addressed to a loopback name (or the --host it binds to) and refuses browser requests from other origins with 403, so a web page on another site cannot use it as its API; --allow-any-host switches that off. The full reference is skills/jev-style/reference.md.

Hosted API

https://api.jevstyle.com (the only documented entry point) serves jev-style-2b-decision-v4 for text only, with an 8,192-token context and inputs of up to 8,000 tokens. There is currently no fixed per-account quota (no daily or per-minute cap); capacity is shared, so when the server is busy it answers 429 with a Retry-After header, and clients should retry after that delay. An input that is too long gets 413 input_too_large there (the local server answers 400 max_tokens_exceeded). This Python client works against it with JevStyle(base_url="https://api.jevstyle.com", api_key=...). The hosted API is a separate service from local inference: check the service status and current limits or the live health endpoint when you need it. A health check is a point-in-time readiness signal, not an end-to-end inference guarantee.

How it works

Each release is a fine-tune of a Qwen3.5 base with every weight trained (the 2B v4 keeps Qwen3.5-2B's vision tower frozen and drops the multi-token-prediction head). A request is rendered as plain text, with no chat template: the state first, then each question with its options and a verdict slot after every option. The model reads the state once; each question is then scored in its own block over that shared state, so one question's options do not change another question's answer. An option's score is logit(" yes") − logit(" no") at its slot. There is no answer-letter cap, a question can have up to 255 options, and the model can only answer with an option it was given, never with free text. Probabilities are softmax(scores / T), with T fitted on held-out calibration rows and never on benchmark items: one global temperature for the v3 releases, one per question type for the 2B v4. The v3 releases were trained on public datasets and generators; the 2B v4 adds soft labels distilled from TypeSafe Jev 1.14 through its public API. No Jev weights or code are used. Each model card gives the full recipe, the data with its licences, and the input format: 2B v4, 0.8B v3.

Limitations

• Hosted Jev is more accurate on every set where its number is published (JevBench public items: 86.6 % against 74.9 % for the 2B v4). Treat these models as small local options, not replacements.
• With enough labels, a classic classifier can win. On the six-intent BANKING77 comparison (240 fixed test requests), TF–IDF + logistic regression got 229 right and the 0.8B 224, and the acceptance rule chosen on the development split exceeded its error budget on the test split. Jev-Style earns its place when you have no labels, several questions per input, images, long inputs, or need probabilities you can threshold. Measure your own labels, calibration, accepted-request errors and end-to-end latency before replacing an existing route.
• JevBench numbers are self-runs with the official harness (the Decision Index numbers are the public board's own), and several benchmarks' official train splits are in the 2B v4's training data (its card lists them), so those results are not zero-shot.
• Images need the 2B v4 PyTorch build. The MLX and GGUF builds hold the language model only. Stock llama.cpp, Ollama, LM Studio and mlx_lm.generate cannot produce the decision scores.
• Calibration is a temperature. It cannot reorder confidences. Recipe thresholds are illustrative, the bundled tickets are not a benchmark, and the 2B was trained on a reduced data pool (60M tokens) with no separate budget for the question and its options.
• Speed on a Mac is hundreds of milliseconds, not tens: about 0.15 to 0.2 s for a short request to the 0.8B. The 2B MLX runtimes need mlx-lm 0.31.3 exactly, so give jev-style its own environment if other MLX tools need a newer mlx-lm. cascade-9b needs about 32 GB of GPU memory.
• The hosted API reads text only, with inputs of up to 8,000 tokens; images and 32,768-token inputs need the 2B v4 running locally.
• The guard is a second line of defence, not a sandbox: 77.6 % agreement on 49 hand-labelled calls, measured with the 0.8B v3; the 2B releases were not measured on that set.

Used by and mentioned in

• The vLLM Semantic Router blog post on System One Auto (2026-10-10) cites Jev-Style's cascade commit as a public precedent for confidence cascades.
• eullm uses the 0.8B v3 GGUF as the decision model behind its auto model routing (setup and benchmark).
• On the Hugging Face Hub, jev-style is a registered model library: the main (PyTorch) model pages show its install snippet under "Use this model" (huggingface.js #2550).
• Listed in yibie/awesome-jev, cobanov/awesome-jev and AnotiaWang/awesome-decision-models, and in the official MCP Registry.

Using Jev-Style somewhere? Open a workflow issue and it will be added here.

Roadmap (ideas, no dates)

Checked items have shipped; unchecked items are proposals, not commitments.

• 0.8B and 2B releases with pinned PyTorch, MLX and GGUF builds; pip install "jev-style[torch]" or "jev-style[mlx]"
• Local server with Playground and demos, Python client, CLI
• Six agent skills, Claude Code guard, MCP server in the official registry
• jev-style eval, engine comparison with paired bootstrap, confidence cascades, CUDA graphs
• Images and 32,768-token inputs (2B v4)
• Proposed: fine-tune a release on your own labelled JSONL, from the same file jev-style eval reads
• Proposed: images in the MLX build
• Proposed: listings in more local model runners

Ideas and requests go in Discussions. The dated history is in the changelog.

Repository layout

jev_style/           server, client, CLI, guard, MCP server, eval; web/ = Playground + demosskills/              six agent skills (npx skills add lawrence3699/jev-style)plugins/             Claude Code guard plugin (the root .claude-plugin/ is the marketplace)examples/            labelled example data for jev-style eval; recipes/; cascade/ = example cascade filesscripts/             cascade_smoke.py: a two-tier cascade on fake engines, prints one routed responsespace/               source of the 0.8B Hugging Face Spacespace-2b/            source of the 2B v4 Hugging Face Spacetests/               pytest suite, runs on the fake engine (no model download)

Contributing and development

Share a minimal failure case or show a workflow in GitHub Issues; ask usage questions and discuss ideas in GitHub Discussions. Please remove private data and credentials. The contribution guide has setup, test commands and small starting tasks; report a security problem privately as described in SECURITY.md.

git clone https://github.com/lawrence3699/jev-style && cd jev-styleuv venv && uv pip install -e ".[dev]"uv run pytest -q                # fake engine: no weights, no GPUuv run jev-style serve --fake   # UI work without the model

Citation

If Jev-Style is useful in your work, cite the repository (GitHub's "Cite this repository" uses CITATION.cff); each model card has its own entry for the weights.

@software{yan2026jevstyle,  author  = {Yan, Chaoliang},  title   = {Jev-Style: small calibrated decision models for AI agents, with a local server, agent skills, MCP tools and a Claude Code guard},  year    = {2026},  url     = {https://github.com/lawrence3699/jev-style},  license = {Apache-2.0}}

About the project

Identity and training sources. JevStyle is maintained by Chaoliang Yan (about the project). It is an independent project with no TypeSafe affiliation or endorsement. The 2B v4 model was trained in part on soft labels distilled from TypeSafe Jev 1.14 through its public API, as disclosed in its model card. The package does not bundle raw teacher-response datasets; that packaging detail does not negate the use of teacher outputs in training. Training sources differ by release; consult each model card.

Local inference and hosted services are separate. Local inference runs the downloaded JevStyle weights on your machine and needs no hosted API key, TypeSafe or OpenRouter account, or agent simulation; the local installation guide covers text, images and backend selection. A third-party Jev skill that simulates decisions with an agent or requires TypeSafe/OpenRouter credentials is a different execution path. The hosted API, the Hugging Face demos and the local server are separate services: check the service status and current limits or the live health endpoint when you need the hosted API. A health check is a point-in-time readiness signal, not an end-to-end inference guarantee. The hosted API's model, context, input limit, retry rule and error shapes are under Hosted API. Published benchmark pages, with their scope, protocols and limitations, are at jevstyle.com/benchmarks.

Acknowledgements and license. Code: Apache-2.0 (LICENSE). The weights are Apache-2.0 fine-tunes of Qwen3.5-2B and Qwen3.5-0.8B; the NOTICE in each model repository lists the changes. The typed-question convention follows Laya. cascade-9b and jevk5: tiers use JevK5-9B by alibiserikbay and its runtime allebee/jevk5, both Apache-2.0; neither is included in this package. Per its model card, part of JevK5-9B's training labels came from OpenAI's GPT-6 Luna under OpenAI's terms.

Not affiliated with, endorsed by or connected to TypeSafe or Jev. "Jev-Style" describes the kind of model: a small typed-decision model in a similar style. This Python package does not distribute TypeSafe model weights, TypeSafe code, or raw teacher API-response datasets; that statement describes the package contents, not the model's training inputs (see above). Not affiliated with Alibaba Cloud or the Qwen team, the Laya authors or the JevK5 author.