Jeff-Qwen3.5-2B

Community preview: v1.2 (1 October 2026). v1.2 is more honest, not smarter: it is trained on the same cleaned data as Jeff-Qwen3.5-0.8B v1.2, with overlaps with our test sets and answer-order and answer-length shortcuts removed. Its benchmark score is level (81.7% against 82.0%), its calibration is better (0.021 against 0.026), and its long-document score drops from 85.8% to 65.6%, because v1.1's figure was inflated by leaked test documents. Details: What changed in v1.2. v1.1 stays available as revision v1.1. The nine LoRA adapters and the comparison with Qwen3.8-27B are for the 0.8B; see its card and jeffhub.ai. A stable long-term-support base (v1.3) is due in about 36 hours.

The Jeff models are fine-tunes of Qwen3.5 and Gemma 4 for zero-shot classification: small, fast decision models you slot into your code. You describe a situation and list the options in plain words; Jeff returns a calibrated probability for each option from a single forward pass. No generated text, no parsing: this model takes about 24 ms per decision on an RTX PRO 6000 and 60 ms on an Apple M4 Max (MLX).

Zero-shot means the options can be anything: support queues, user intents, moderation labels, voice commands, game moves. Your categories don't need to appear in the training data; you describe them, and Jeff picks.

What it is, and what it isn't. These are very small models. They make extremely fast, well-calibrated judgement calls between options, and they slot easily into your local code. On benchmarks they approach, and sometimes beat, Jev; but at this size their reasoning won't match Jev's, which runs on a much larger model. If zero-shot accuracy isn't good enough for your purposes, a short fine-tune on your own examples takes you much further: our voice-navigation fine-tune moved held-out accuracy from 31.7% to 95.8% in under half an hour on one GPU.

Built entirely on local hardware. Training on one RTX PRO 6000 workstation GPU (the 0.8B trains in about 2 hours, the 2B in about 3.5), all synthetic training data of the base model written by an open model (Qwen3.8-Flash-Next) on two DGX Sparks, testing on a MacBook. No cloud GPUs, and no closed model was used as a teacher for the base model; a closed model was used only to spot-check the quality of a sample of the synthetic data. Some public data sets in the mix contain text their authors generated with closed models (for example the model responses in RAGTruth).

Independent project. Jeff uses the same request format as Jev, but it is not affiliated with or endorsed by TypeSafe, the makers of Jev. Our training code starts from the open-source AutoJev recipe.

Model	Base model	Base model's panel accuracy (untrained)	Jeff's panel accuracy	Calibration error (ECE)	Decision time (RTX PRO 6000)
Jeff-Qwen3.5-0.8B (v1.2)	Qwen3.5-0.8B	45.3%	78.7%	0.028	22 ms
Jeff-Qwen3.5-2B v1.2 (this model)	Qwen3.5-2B	46.5%	81.7%	0.021	24 ms
Jeff-Qwen3.5-2B v1.1 (revision v1.1)	Qwen3.5-2B	46.5%	82.0%	0.026	24 ms
Jeff-Gemma4-E2B	Gemma 4 E2B	62.5%	81.6%	0.031	29 ms
Jev (published)	—	—	83.0%	≈0.06 (average of its per-benchmark figures)	212 ms per call over the API (Doom harness)

What changed in v1.2

v1.2 is a new training run of the 2B on the cleaned v1.2 data (284,747 questions, one epoch, final checkpoint, one fitted temperature), the same data as Jeff-Qwen3.5-0.8B v1.2. Its card lists the data changes in full: leaks removed (every MAUD question, among others), answer-order and answer-length shortcuts removed, and voice navigation moved out of the base into its own adapter.

Test	Questions	v1.1	v1.2	Note
Benchmark panel (five benchmarks)	4,599	82.0%	81.7%
Calibration error on the panel (ECE)	4,599	0.026	0.021	lower is better
Long lists v2 (20 to 254 options; a new test with no message seen in training)	1,886	not measured	93.6%
Long lists v1	2,400	95.2%	95.0%	v1 shares messages with the training data; v2 replaces it
Long documents (ContractNLI, CUAD, MAUD, ConditionalQA)	2,009	85.8%	65.6%	v1.1's score was inflated by leaked MAUD documents
Voice navigation (held-out spoken commands from our app)	3,324	95.8%	91.4%	now zero-shot: the voice data moved to the nav adapter
JevBench hard tier	105	57.1%	57.1%

Per benchmark (v1.1 → v1.2): BBH 68.7 → 66.4, Financial PhraseBank 94.7 → 95.6, JudgeBench 59.4 → 62.0, RAGTruth 87.7 → 85.5, WinoGrande 78.8 → 80.7. We read v1.2 as level with v1.1 on the panel and better calibrated. If you rely on v1.1 for long contracts, expect v1.2's figure, not v1.1's.

What changed in v1.1 (the Qwen models)

• Long lists. The v1.0 models never picked an option past the 26th, because no training question had more than 19 options (thanks to @puhuk for the report, #1). v1.1 adds 32,000 training questions with 20 to 254 options and accepts up to 254. On our long-list test (lists of 20 to 254 items) Jeff-Qwen3.5-0.8B goes from 40.3% to 94.7%, and Jeff-Qwen3.5-2B scores 95.2%.
• Final checkpoint. v1.1 publishes the checkpoint at the end of the single epoch. Choosing the checkpoint with the lowest development loss, as v1.0 did, picked an early, less settled checkpoint in the new runs and cost 1–2 points.
• Calibration is better: calibration error 0.049 → 0.021 for the 0.8B and 0.028 → 0.026 for the 2B. JevBench's hard tier: 47.6% → 46.7% for the 0.8B, 53.3% → 57.1% for the 2B.
• Benchmarks: the 0.8B is unchanged at 79.1%. The 2B goes from 83.1% to 82.0%, mostly on JudgeBench (64.6% → 59.4%), a small benchmark where models of this size sit close to chance; so the 2B no longer edges past Jev's published 83.0%.
• Training data also adds the training splits of MASSIVE and CLINC150 (as full intent lists), so results on those two data sets are no longer zero-shot, and more synthetic data from our local teacher.
• v1.0 is still available: add --revision v1.0 to hf download. The game results below were measured with v1.0.
• Jeff-Gemma4-E2B was not retrained and stays at v1.0.

What it does

{
  "model": "jeff-latest",
  "state": {"voice_transcript": "open the engagment leter", "current_screen": "Deal overview"},
  "questions": {
    "intent": {
      "type": "choice",
      "instructions": "Which of these does the user want?",
      "criteria": {"1": "Engagement letter", "2": "Inbox", "3": "Deal settings"}
    }
  }
}

The answer is a probability per option ({"1": 0.94, "2": 0.03, "3": 0.03}), the chosen option and a confidence. Three question types: choice (pick one of up to 254 options with the Qwen models, 26 with Jeff-Gemma4-E2B), noul (yes/no, returned as a probability) and score (a point on a scale you describe). Several independent questions in one request are answered together.

Benchmarks

4,599 questions from five public benchmarks, plus JevBench's public hard tier (105 items, scored separately). The Qwen models are v1.2, Jeff-Gemma4-E2B v1.0:

Benchmark	Qwen3.5-0.8B untrained	Jeff-Qwen3.5-0.8B v1.2	Qwen3.5-2B untrained	Jeff-Qwen3.5-2B v1.2	Gemma 4 E2B untrained	Jeff-Gemma4-E2B	Jev (published)	AutoJev-27B (published)
Overall (5 benchmarks)	45.3	78.7	46.5	81.7	62.5	81.6	83.0	84.9
BBH	39.5	63.2	46.0	66.4	51.3	66.4	94.3	82.8
Financial PhraseBank	36.0	96.3	53.4	95.6	86.0	96.1	77.0	84.2
JudgeBench	56.6	60.9	57.4	62.0	46.9	60.6	78.6	78.9
RAGTruth	49.1	85.5	35.9	85.5	63.8	87.4	77.3	88.9
WinoGrande	49.2	68.7	52.2	80.7	51.0	77.4	90.7	83.3
JevBench hard (separate)	36.2	44.8	45.7	57.1	41.0	48.6	73.3	70.3

Bold: the winner of Jeff against Jev in each row: a Jeff score above Jev's published figure, or Jev's figure where it beats every Jeff model. Bold italic: AutoJev-27B where it is the best of all models in the row; it is shown for reference, since the head-to-head comparison is with Jev. The published Jev and AutoJev figures were measured on a different sample of the same benchmarks. Jeff's overall score comes from classification and grounding (Financial PhraseBank, RAGTruth), where it matches or beats the large models; on the reasoning-heavy benchmarks (BBH, JudgeBench, JevBench) it stays well below them, as you would expect at this size.

Games: a zero-shot test

To test zero-shot performance on tasks unlike anything in the benchmarks, we had Jeff play three games. Games aren't the ideal zero-shot test, since a game's state isn't typical unstructured data; but they are a common, and fun, way to test a System 1 model. Each turn, the code describes the situation and the legal moves in words, and the model picks one. The options state what each move leads to (Frogger: "you would be hit by a car and lose a life"; Doom: "the nearest monster is a little to your left"), but never which move is right. Each result is 20 episodes, seed 1234; ▶ opens a video of the run's first episode. The games were measured with v1.0 of the Jeff models.

Jeff-Qwen3.5-0.8B playing, zero-shot (the bold row in the table below):

Doom

Frogger

Pac-Man

Model	Doom, kills (monster's direction in words)		Frogger, crossings (consequences)		Pac-Man, pellets of 98 (consequences)
Random moves	−0.05		0		11.2
Hand-coded rule bot	6.55	▶	10.25	▶	94.1	▶
Qwen3.5-0.8B, untrained	5.0	▶	1.0	▶	25.8	▶
Jeff-Qwen3.5-0.8B	6.55	▶	10.3	▶	57.0	▶
Qwen3.5-2B, untrained	0.55	▶	0.05	▶	72.1	▶
Jeff-Qwen3.5-2B	−0.9	▶	6.0	▶	41.2	▶
Gemma 4 E2B, untrained	−0.55	▶	0	▶	3.2	▶
Jeff-Gemma4-E2B	0.55	▶	0.15	▶	53.2	▶
Jev (published, Doom)	6.55, told the aiming rule; −0.60 without it		—		—

Jeff-0.8B decides in 29–49 ms per move on an M4 Max; Jev's published Doom run took 212 ms per call over its API. The two times were not measured on the same hardware.

Speed and size

Median time per decision over the same 200 benchmark questions (about 200 input tokens each), one question at a time, from raw text to probabilities:

Model	Parameters	Weights (16-bit)	NVIDIA RTX PRO 6000	Apple M4 Max (MLX)	CPU (32 threads)
Jeff-Qwen3.5-0.8B	0.8B	1.7 GB	22 ms	28 ms	463 ms
Jeff-Qwen3.5-2B	2B	4.2 GB	24 ms	60 ms	708 ms
Jeff-Gemma4-E2B	2B effective (4.6B stored)	9.3 GB	29 ms	— (MLX runs Qwen only)	1.0 s
AutoJev-27B	27B	~54 GB	not published	—	—
Jev	not disclosed	API only	114–212 ms per call in published Doom runs, including the network

How to use

git clone https://github.com/firelex/jeff && cd jeff
uv sync --no-default-groups   # serving only; add --extra cuda on NVIDIA GPUs (fast kernels), --extra mac on Apple silicon
uv run --no-default-groups hf download mstrasser/Jeff-Qwen3.5-2B --local-dir Jeff-Qwen3.5-2B
JEFF_CHECKPOINT=Jeff-Qwen3.5-2B PORT=8765 uv run --no-default-groups jeff-serve                        # NVIDIA or CPU
JEFF_BACKEND=mlx JEFF_CHECKPOINT=Jeff-Qwen3.5-2B PORT=8765 uv run --no-default-groups --extra mac jeff-serve       # Apple silicon
curl -s localhost:8765/v1/systemone -H 'content-type: application/json' -d @request.json

• Reason in code, decide with Jeff. State each option's consequence; don't ask Jeff to forecast.
• Use short option keys and descriptive text: {"1": "Engagement letter"}, not long IDs, which cost time and add nothing.
• Fine-tune for your domain. A voice-navigation fine-tune on ~11k app-specific examples took held-out accuracy from 31.7% to 95.8% in one epoch.

Training

• Recipe: full-weight supervised fine-tuning, one epoch (learning rate 5e-6 for the 0.8B, 1e-5 for the 2B), batches of 256, cross-entropy over the option letters, then one scalar temperature fitted for calibration. From v1.1 we publish the checkpoint at the end of the epoch; the benchmark panel is never used for selection or tuning.
• Data (285k questions in v1.2; 322k in v1.1; 271k in v1.0): v1.2 is the cleaned mix described on the 0.8B's card. The kinds of sources are those of v1.1: public datasets converted to decisions (entailment, QA, sentiment, safety, fact verification and more), the training splits of WinoGrande (40k) and RAGTruth (15k), long real documents (ContractNLI, CUAD, ConditionalQA), 10k code-built probability questions with exact answers, ~50k synthetic questions written and checked by a local teacher model (Qwen3.8-Flash-Next on DGX Sparks), hijack attempts planted in 3% of questions, "none of these" options added to 5%, and (v1.1) 32k long-list questions with 20–254 options, from code and from the MASSIVE and CLINC150 training splits.
• Leak filter: every training question is checked against the benchmark panel and JevBench; 50 near-duplicates were removed.
• Disclosure: at least half of each training family is written in the benchmark panel's layout conventions (how states, questions and options are formatted). No panel item is used in training, but matching the format helps the score.
• Hardware: one RTX PRO 6000 (96 GB) for training, DGX Sparks for the teacher, and an M4 Max for local tests.

Caveats

• Option limits. The Qwen models are trained on lists of 20 to 254 options and accept up to 254. Jeff-Gemma4-E2B is still v1.0: its training never had more than 19 options, so it only picks reliably among up to 26, and the server refuses longer lists for it; shortlist first. Thanks to @puhuk for the report (#1).
• Small models don't reason. Expect fast, calibrated choices between the options you describe, not multi-step reasoning. At 0.8B–2B parameters this holds for every model, not just Jeff.
• Jeff-2B is a weaker game player than Jeff-0.8B. The untrained 2B already appears more risk-averse than the untrained 0.8B (in Doom it prefers turning away from the nearest monster; in Pac-Man it survives much longer but hesitates), and our training seems to have made that worse: Jeff-2B reverses direction in Pac-Man 3.5× as often as the untrained 2B. This needs more investigation.
• Benchmark scores don't predict game play. The untrained Gemma 4 E2B beats the untrained Qwen models on the benchmarks (62.5% against 45–47%) yet plays the games worst: it makes the right move most of the time but not reliably, and in a real-time loop the occasional wrong call compounds. Training fixed its Pac-Man (3.2 → 53.2 pellets) but not its Doom or Frogger.
• Prompts matter. The game results depend on options that state consequences in words. Jev's own Doom prompt (a raw bearing number plus an aiming rule) does not work for any of our models, trained or untrained.
• Different samples. The published Jev and AutoJev numbers were measured on a different sample of the benchmarks.
• English and text only. Calibration is fitted on our development data; recalibrate for a very different domain.

Licence and data

Weights: Apache 2.0 (from Qwen3.5-2B by the Qwen team, Alibaba Cloud). The training data mixes datasets under various licences, listed with their sources in docs/data-sources.md. We release the weights and code, not the training data; some sources are share-alike (CC BY-SA).