Three multimodal releases landed in the same month, and they share almost nothing on the surface. A 4B fine-tune ranked #1 of 91 on a decision benchmark. JetBrains merged two 27B checkpoints into an image-text model without running any training. And an OCR checkpoint with almost no documentation quietly did its job. The thread connecting them matters more than any single release: vision-language capability is no longer gated on frontier-scale training runs.

Three releases, one pattern

ImaJev-4B is a fine-tune. Qwen3.8-3.6-27B-blend is a merge. TeleOCR is a specialist. Fine-tune, merge, specialize. Those are the three affordable paths to multimodal in 2026, and this cluster tells you exactly what each one costs and where it breaks.

The expensive path, pre-training a frontier multimodal model from scratch, still exists. It's just no longer the only road. If you have a GPU, a few thousand dollars, or even just a merge script, you're in the game.

ImaJev-4B: decisions from photos and JSON

ImaJev started with an observation from business process consulting. Process maps for refunds, returns, and customer support have decision nodes everywhere, and almost all of them are staffed by humans. Encoding ambiguity in code is hard, so condition nodes stay manual. The base model, Jev, handled text decisions well but had no vision support. The question was whether one model could do both.

The answer is a compact architecture: LoRA plus a small decision head on Qwen3.5-4B. You feed it text or a JSON record, up to two photos, and a closed question. It returns a probability for each option plus an "unknown" class. All in one forward pass. No token generation, no autoregressive decoding. For a decision-service use case, that latency profile is the difference between a synchronous API call and a streaming job.

4B parameters means the whole thing runs on a Mac with MLX or on a single GPU. No cluster, no cloud reservation. The training cost was roughly $1,200 in rented GPU time. That's a small-business budget, not an ML-team line item.

Metric	Value	Practical meaning
Model size	4B parameters	Runs on a Mac (MLX) or one GPU
Training budget	~$1,200 rented GPU	Within reach of a solo consultant
JevBench rank	#1 of 91	Top of the weighted leaderboard
JevBench score	67.37 vs 63.29	4-point lead over Jev 1.13.0
Accuracy-only rank	#3	Weighted score flatters it slightly
DecisionBench rank	#3 of 56	Ahead of GPT-5.6 Luna and DeepSeek V4.1

The collapse that taught the real lesson

My first fine-tune was a textbook mistake, and it failed loudly. I trained on about 500k short decisions, the kind of easy examples you can synthesize in bulk. JevBench hard went from 64.9 down to 42.3. A 22-point drop. The model had learned to pattern-match. It stopped reasoning and became a very expensive lookup table.

The fix took two weeks and one key insight: difficulty and agreement. I generated hard questions with open-weight models and kept only the ones where two separate models agreed on the answer. Volume didn't matter. The filter did. That single change brought the score back, and the current ImaJev checkpoint ended up at 67.37 on JevBench v1.4.2.2, ahead of the base model that collapsed.

The lesson transfers to any decision-model project. Bulk synthetic data does not teach reasoning. Hard, adjudicated examples do. If your fine-tune data doesn't force the model to think, the model will find the shortcut instead.

Reading the benchmark results carefully

The headline is "#1 of 91," and you should read the fine print before acting on it. JevBench's score weights accuracy, calibration, speed, and cost equally. On accuracy alone, ImaJev sits at #3. Still strong for a 4B model, but the top rank overstates raw intelligence.

The calibration story is the more interesting one anyway. When ImaJev says 90%, it's usually right, and it outputs "can't tell" instead of forcing a guess. For decision support, honesty under uncertainty is worth as much as raw accuracy. The same week, DecisionBench placed it #3 of 56, ahead of GPT-5.6 Luna and DeepSeek V4.1.

Quick Take: A 4B model hit #1 on a 91-model decision board by pairing top-3 accuracy with the calibration, speed, and cost that bigger models can't match.

Merging instead of training

JetBrains took a different route entirely. Qwen3.8-3.6-27B-blend is a 50/50 linear interpolation of the Qwen3.6-27B and Qwen3.8-27B checkpoints, accumulated in float32. No training, no new data. The configuration, tokenizer, processor, and chat template all come from Qwen3.8-27B. What changes are the weights, and apparently that's enough.

On JetBrains' internal 100-task coding benchmark, the blend completes tasks with fewer output tokens than either parent. That's a token-efficiency gain, which in practice means lower latency and lower cost per request. For local tooling, that's often the deciding factor.

A 27B model is still not small, but the available formats make it portable: BF16, GGUF, and MLX 4-bit. The 4-bit versions change the hardware story. MLX runs on Apple Silicon, GGUF fits on a 24GB consumer card like an RTX 4090. 645 downloads last month is modest, but the technique matters more than this specific model.

The catch is trust. A merge inherits both parents' quirks, and a 50/50 blend can behave like neither. The merge manifest records the exact source revisions, which is honest engineering, but it doesn't tell you how the model behaves on your documents. Run your own eval before you rely on it.

OCR, the unglamorous specialist

TeleOCR's model card is nearly empty. Name, a link, nothing to evaluate. That emptiness is typical of the OCR niche, where checkpoints ship as utilities rather than research artifacts. Don't let the thin documentation fool you though. In multimodal pipelines, OCR is where most real-world accuracy gets won or lost.

Think about what ImaJev actually needs to do: two photos in, a business decision out. For that to work, the text in those photos has to be extracted cleanly. Receipts, invoices, screenshots, ID cards. A general VLM can read these well enough to describe them, but a dedicated OCR model lives and dies by exact character extraction. The practical pattern is routing: OCR the document first, hand clean text to the reasoning model, and skip pixel-level reasoning in the big model entirely. Cheaper and more accurate than asking one model to do both jobs.

Which path fits which problem

	ImaJev-4B	Qwen blend	TeleOCR
Method	LoRA + decision head fine-tune	50/50 checkpoint merge	Specialist OCR training
Input	Text/JSON + up to 2 photos	Image + text	Document images
Output	Probabilities + "unknown"	Generated text	Extracted text
Hardware	Mac (MLX) or 1 GPU	24GB GPU or Apple Silicon (MLX 4-bit)	Varies
License	Apache-2.0	Apache-2.0	Not documented
Best for	Decision automation with ambiguity	Cheap local multimodal inference	Receipts and dense documents

None of these paths needed a training budget a small team can't afford. ImaJev was $1,200 of rented GPUs. The blend cost a merge run and some careful evaluation. OCR specialists are routinely trained on modest datasets. The expensive path is no longer the default.

Common Pitfalls

Five mistakes show up in this cluster, and each one has a body.

Fine-tuning on bulk easy data. ImaJev's first run trained on 500k short decisions and JevBench hard dropped 22 points. The model learned to pattern-match. If your fine-tune data never forces reasoning, your model will stop reasoning. Filter for difficulty and agreement instead of maximizing volume.

Trusting weighted leaderboard ranks. JevBench's #1 averages accuracy, calibration, speed, and cost. On accuracy alone the model is #3. Weighted scores mix together things that matter differently to different users. Always check the per-axis numbers.

Merging checkpoints without your own eval. The Qwen blend inherits blind spots from both parents. The merge manifest tells you the math, not the behavior. Test on your actual tasks before deploying.

Using a general VLM for text extraction. If your job is reading receipts or dense documents, a specialist OCR model will beat a generalist on character-level accuracy. Route extraction to OCR and reasoning to the language model.

Forcing a decision instead of allowing "unknown." A model that must pick an option will fabricate confidence. ImaJev's explicit unknown class is exactly what makes its 90% claims believable. Build the same escape hatch into your system.

One Thing to Remember

Calibration and cost are competitive weapons. ImaJev didn't win the leaderboard by being the smartest model. It won because its probabilities mean something, it runs on a laptop-class setup, and it knows when to say "can't tell." The Qwen blend's entire pitch is token efficiency. In multimodal, "good enough and honest about being unsure" is a winning product position.

The Bottom Line

If you're building a document-heavy pipeline, route OCR first and hand the clean text to a small VLM. You'll get better text accuracy and lower cost than asking one frontier model to read everything.

If you're automating business decisions with real ambiguity, follow the ImaJev recipe: fine-tune a small model on hard, agreement-filtered examples, and give it an explicit unknown output. A 4B model can beat models several times its size on the metrics that matter to your users.

If you have no training budget at all, checkpoint merging gives you a workable multimodal model for the cost of a merge run. But treat the blend as unverified until your own eval says otherwise. One thing to watch: merge tooling is improving fast, expect blends like this to become a standard step in local-model workflows within a year.