DeepSeek V4 was released in April 2026, with two models: V4-Pro, 1.6T total parameters, 49B activated; V4-Flash, 284B total parameters, 13B activated. Both natively support 1 million token context, three-tier reasoning modes ranging from fast direct answers to step-by-step reasoning to maximum reasoning effort, and also support function calling. V4-Flash is priced at $0.14 per million input tokens and $0.28 per million output tokens, roughly 36 times cheaper than Claude Opus 5. In terms of capability, V4 lags behind closed-source frontier by about 3 to 6 months, while priced at a fraction of the cost. On August 16, DeepSeek switched to peak/off-peak pricing: V4-Flash peak rose to $0.44/$1.32, off-peak $0.22/$0.66, roughly 1.6 to 4.7 times the original. Even after the increase, V4-Flash remains one of the cheapest reasoning APIs on the market, and Claude Opus 5’s output price is still 19 times V4-Flash’s peak output price.
This isn’t luck from a single version. DeepSeek has been pursuing efficiency since V2. The V2 paper is literally titled “A Strong, Economical, and Efficient Mixture-of-Experts Language Model.” It introduced Multi-head Latent Attention, which compresses the KV cache into a low-dimensional latent vector, and DeepSeekMoE, which uses fine-grained expert segmentation to reduce training cost. V3 continued this line: 671B total parameters but only 37B activated per token. Training cost $5.57 million, with capabilities comparable to GPT-4-class models. V4-Pro’s CSA+HCA hybrid attention compresses the compute for million-token inference to 27% of V3.2, and the KV cache to 10%.
Send an image_url field to this engine and you get a
JSON deserialization error:
Failed to deserialize the JSON body into the target type: messages[0]: unknown variant ‘image_url’, expected ‘text’
The request is blocked before it even reaches the reasoning engine. The API schema has no image input option. The model catalog lists 11 DeepSeek models, all with modality text.
It’s not that DeepSeek can’t do vision. In December 2024 it released a standalone vision model DeepSeek-VL2, scoring 834 on OCRBench, surpassing GPT-4o at release. In July 2026, a DeepSeek researcher briefly published a paper titled “Thinking with Visual Primitives” on GitHub, took it down 4 hours later, the original repo is now 404, leaving only community mirrors and explainer videos. The paper wove bounding boxes and points directly into the reasoning chain, achieving 66.9% on maze navigation tasks, versus about 49-51% for GPT-5.4 over the same period. But the paper was retracted, the weights were never released, and the official statement only said it would be integrated into the base model in the future, with no timeline. VL2’s context window is only 4096 tokens, with no function calling, and barely iterated after release. None of these vision capabilities made it into the commercial API.
Not doing multimodal is a current product choice, not a technical limitation. DeepSeek has a vision research line and frontier papers, and may add native multimodal at some point in the future. When that point is, nobody knows.
On August 12, 2026, Liquid AI released LFM2.5-VL-3B, a 3.1B-parameter vision-language model. The model card has a rare, explicit statement: it recommends the model for single-turn, high-throughput, low-latency tasks such as real-time object detection, batch OCR, and menu/signage translation. It then clearly states it is not recommended for long-context or heavy-reasoning tasks such as visual web design or answering technical questions about blueprints. Most small VLMs try to prove they can do everything. This model card does the opposite, stating what it is not good at.
LFM2.5-VL-3B packs 3.1B parameters into 3.3 GB of memory, running 228 tokens/s on Apple M5 Max, and 20 tokens/s on a Galaxy S26 Ultra phone. In the vendor’s cross-benchmark, screen grounding ScreenSpot-v2 scores 80.7, object grounding RefCOCO 87.9, ToolSandbox 59.5. These numbers lead the 3B class, with some metrics approaching 5B models. All come from Liquid AI’s own cross-benchmark, with no independent reproduction yet. Someone on Reddit specifically posted asking whether there were external benchmarks, and none were found.
This model specializes in perception and doesn’t touch reasoning; DeepSeek V4 specializes in reasoning and doesn’t touch perception. Before DeepSeek natively supports multimodal, these two models can be assembled into a pipeline: LFM2.5-VL-3B does visual perception at the edge, outputting structured text (the door is open, there is a car in the frame), while DeepSeek V4 does reasoning in the cloud (is this the owner’s car or an anomaly? what notification should be triggered?).
After LFM2.5-VL-3B was released, I ran an evaluation on my home garage camera data. The validation set had 280 images, manually labeled for garage door open or closed, where any visible opening counts as open, and only fully closed counts as closed. LFM2.5-VL-3B had never seen any garage footage. Zero-shot, it achieved 94.3% accuracy, with 92.6% recall on open doors. Compared to a small model specifically trained on this task, the gap is about 3 to 5 percentage points. A 3B general VLM trained on zero garage data approached a specialized model on a clear single-question judgment.
But if you switch the task to multi-class object detection, looking for whether there is a car or a person among 13 classes, recall drops to 66.2%, with a 47% false positive rate on empty scenes. A zero-shot VLM is good enough for simple binary judgments, but not yet for fine-grained multi-class detection. This gap doesn’t undermine the value of the division-of-labor architecture, because the perception layer doesn’t need to replace specialized CNNs for everything.
The perception layer makes the first-tier judgment: is the door open? is there a person? This layer is well served by a general VLM zero-shot. When an anomaly is detected, the structured perception result is sent to DeepSeek V4 for reasoning. Perception completes locally in 1.5 seconds, and reasoning in the cloud takes another 1 to 3 seconds. For scenarios that only trigger on anomalies and don’t need per-frame judgment, this latency is acceptable.
The separated architecture has things that native multimodal can’t offer.
First is latency. Perception completes locally in 1.5 seconds, with no network dependency. If you put vision and reasoning in the same cloud-based multimodal model, image upload, visual encoding, and reasoning generation add up to at least 3 to 5 seconds, plus network round-trip. The latency gap between 1.5-second local perception and 3 to 5-second cloud reasoning is a physical constraint, not an optimization opportunity. No software upgrade can eliminate it.
Second is privacy. Perception completes locally, and images never leave the device. Only structured text is uploaded to the cloud. For scenarios like home surveillance, medical documents, and financial forms, not uploading the image itself is an independent safety property, unrelated to model capability.
Third is cost. High-frequency perception runs on a 3B model locally, with zero marginal cost. DeepSeek API is only called when an anomaly is detected. If you ran cloud reasoning on a multimodal large model for every frame, the cost would be several orders of magnitude higher.
The industry is already moving in this direction. In July 2026, viso.ai split the architecture of production-grade vision AI into three layers: a perception model captures the scene, a vision-language model interprets the scene, and an action layer closes the loop. They wrote a key constraint: the perception layer determines the cadence of the entire agent workflow. If detection results arrive at the reasoning layer too slowly or inaccurately, all downstream decisions degrade. Moondream is a 2B-parameter small VLM; some developers combine it with a large LLM, where the large model does planning and Moondream does UI understanding, which is faster and more reliable than directly using the Computer Use API. In January 2026, the vLLM project released Semantic Router, a routing layer between the user and the model. RouteLLM’s data shows that maintaining 95% of quality can reduce cost by 45-85%. Apple Intelligence processes about 85% of requests on-device, 12% via Private Cloud Compute, and 3% via partner clouds. Apple layers by deployment location, but the idea is the same: upgrade on demand, cost-aware routing.
DeepSeek might release V4-VL tomorrow, adding vision to the API. Even so, the three advantages above still hold. The latency gap between 1.5-second local perception and 3 to 5-second cloud reasoning cannot be eliminated by any software upgrade. The privacy property of images not leaving the device doesn’t become less important just because the model gets stronger, and the cost advantage of running high-frequency perception on a 3B model locally doesn’t disappear just because API prices drop. One model doing both vision and reasoning doesn’t mean they should be done in the same layer.
There’s still a gap between concept viability and production readiness. The cascade architecture needs a confidence signal to decide when to escalate from the perception layer to the reasoning layer. LFM2.5-VL-3B outputs natural language, not probability scores. How to extract a reliable confidence signal from natural language is the piece that needs human work in migrating this architecture from LLM cascades to VLM cascades. The TMLS model routing report points this out: the core obstacle to migrating the cascade methodology already validated in the LLM domain to multimodal is the calibration of the confidence signal. This is an engineering problem to solve, not one that disappears by waiting for model capability to improve.
When DeepSeek will add vision is unknown. It has a vision research line, frontier papers, and technical reserves. Maybe next version, maybe next year, maybe longer. The value of the separated architecture doesn’t depend on that timing. Before DeepSeek grows eyes, using a 3B model for perception and DeepSeek for reasoning is a workaround that runs today. The engineering experience accumulated after getting it running—how to bridge perception output to reasoning input, how to calibrate confidence, how to design escalation logic—won’t go to waste when native multimodal arrives. The advantages of division of labor at the physical layer outlast any single company’s product line.