DeepSeek V4 Flash Vision Edition Tested: Blazing Speed, Disappointing Design Chops

Giving a Text Model "Eyes": DeepSeek's Vision Edition Arrives

DeepSeek recently opened up the vision version of V4 Flash via API, sparking widespread interest in the local deployment community. As a speed-focused model, how does it perform on a dual-Spark setup? Is the addition of vision capabilities a genuine upgrade, or just window dressing? A Bilibili content creator ran a full evaluation of this experimental model and benchmarked it directly against GLM 5.3 Flash, QWEN 3.8 Flash, and other competitors.

The conclusion is telling: this model is extremely fast with solid agentic capabilities, but underwhelming on visual design tasks. The tester put it perfectly — "They gave a text model eyes, but didn't build a model that was born with eyes." That observation cuts right to the core tension in today's multimodal model adaptation work.

Why Benchmark Rankings Don't Match Real-World Feel

According to Artificial Analysis evaluations, this vision model is ranked slightly lower because it trades away some visual capability. Yet on the Agentic Index, it performs remarkably well — ranking higher than 27B models, whereas the previous 0731 version sat at just 48th place.

Agentic capability refers to an LLM's ability to autonomously plan, invoke tools, and execute multi-step tasks — not simply answer single-turn questions. A capable agent model can understand complex instructions, break tasks into sub-steps, call external tools like browsers, code executors, and APIs, and dynamically adjust strategy based on intermediate results. The Agentic Index is a specialized benchmark measuring exactly this, evaluating tool-call accuracy, multi-step reasoning success rates, and task completion efficiency. This metric matters because traditional benchmarks like MMLU and HumanEval primarily test knowledge and coding ability — they can't reflect how a model actually performs in real-world automated workflows.

This ranking discrepancy reflects a real limitation: a single composite benchmark score rarely tells the full story about a model's capabilities. The tester noted that DeepSeek V4.5 is reportedly due sometime in September, and what's being tested here is still an experimental version, not an official release.

One detail worth noting: how vision capability is implemented matters a great deal. The tester observed that the model doesn't invoke any external tools (like SIM or tool calls) when processing images — the model itself directly "perceives" the image content, something its previous text-only version couldn't do. The level of detail invested in the reasoning process suggests DeepSeek has integrated a fairly robust image encoding module. This means visual information isn't piped in via an external channel; it's processed directly during the model's forward inference pass, with image features converted through a visual encoder and fed directly into the model's internal attention computation.

Speed Benchmarks: Who Wins on Dual Spark?

Speed is the core selling point of DeepSeek V4 Flash. On a dual-Spark setup, the key numbers from testing are:

• DeepSeek V4 Flash: Peak ~53 tokens/sec (FP8 quantization)
• GLM 5.3 Flash: Peak ~46 tokens/sec
• QWEN 3.8 under TP4 tensor parallelism: Peak 100–115 tokens/sec

A few key technical concepts are worth unpacking here. A dual-Spark system means using two high-performance GPUs working in tandem for local deployment. Tensor Parallelism (TP) is one of the core strategies in distributed inference — it splits the model's weight matrices along specific dimensions across multiple GPUs, with each GPU handling a portion of the computation and results aggregated via high-speed interconnects like NVLink. TP2 means splitting across 2 GPUs; TP4 means splitting across 4. The advantage of tensor parallelism is significantly reduced VRAM pressure per card and higher throughput, but its efficiency is highly dependent on inter-GPU communication bandwidth — if GPUs are connected via PCIe rather than NVLink, communication bottlenecks can negate the speed gains from parallelism. This explains why QWEN 3.8 achieves higher token speeds under TP4.

The tester emphasized that all three models perform solidly in single-stream and dual-stream scenarios; differences in scalability only become apparent under higher concurrency. He was running the official FP8 quantized version provided by DeepSeek. FP8 (8-bit floating point) quantization compresses model parameters from the typical FP16 or FP32 precision down to an 8-bit floating-point representation. Unlike traditional INT8 integer quantization, FP8 retains exponent bits, giving it more flexibility in representing very large or very small values and better preserving the model's original precision. In local deployment scenarios, FP8 quantization can reduce VRAM usage by roughly half while boosting inference throughput by 30–50%, typically with under 1% accuracy loss. The tester noted that RTX 6000 or even 3090 cards could be used to replicate the setup with different quantization configurations.

One important point: a fast model doesn't necessarily mean tasks get completed faster. The tester repeatedly emphasized that a smarter model may actually complete tasks more quickly through more precise reasoning, so "peak token speed" needs to be interpreted alongside actual task completion efficiency. This is especially true in agent scenarios — a high-speed model that requires multiple retries may take far longer in total than a slightly slower model that gets it right the first time.

Agentic Capabilities: Where DeepSeek V4 Flash Really Shines

In the actual testing framework, DeepSeek V4 Flash demonstrated a complete agentic workflow — it autonomously deployed a browser and used it to fetch real-time world news. Latency performance throughout was excellent, with very low Time to First Token (TTFT). TTFT is a key metric for LLM inference responsiveness, measuring the time between a user sending a request and the model outputting its first token. This latency is primarily driven by the prefill phase, where the model performs a forward pass over the entire input sequence, plus system-level scheduling and network transmission overhead. For agent scenarios, low TTFT is critical — agents typically need to call the model multiple times, and each round's first-token latency compounds, directly impacting the total time to complete a task chain. In vision models, the additional computational overhead of image encoding often increases TTFT, so DeepSeek's vision version maintaining low latency speaks to the efficiency of its image encoding module.

Agentic index comparisons across models are quite close:

Model	Agentic Score
DeepSeek V4 Flash	52
GLM 5.2 Flash	53
GLM 5.3 Flash	57
QWEN 3.8 Flash	56

The tester candidly acknowledged that DeepSeek V4 Flash can handle roughly 90% of his daily workload, particularly from an agentic perspective — "It's fast enough and smart enough to handle the vast majority of daily tasks as a primary model." For users who aren't building entirely new projects but just running agent tasks with local models, DeepSeek's speed advantage is genuinely compelling.

Weaknesses Exposed: Mediocre Design Output and High Hallucination Rates

However, the addition of vision capabilities hasn't translated into design ability. The tester tried giving the model a web design task and the result was "terrible" — the output was chaotic and nearly crashed his computer.

His analysis cuts deep:

"Giving a model vision doesn't automatically mean it can do design. But if you train a model with vision from scratch, that's a completely different story."

This reveals a fundamental difference between "vision bolted on later" and "natively multimodal training." Today's multimodal large models take two main technical approaches. The first is late fusion: start with a strong pre-trained text model, then attach a pre-trained visual encoder (like CLIP's ViT) at the input, map image features to the text model's embedding space via a lightweight projection layer, and fine-tune with limited multimodal data. This approach is low-cost and fast to iterate, but visual information is essentially an "add-on" input — the model's core reasoning logic still revolves around text. The second is early fusion (native multimodal training): feed interleaved image-text data from the pre-training stage, letting the model learn joint visual-linguistic representations from the ground up. This is far more expensive to train, but the model's understanding of visual information is deeper and more internalized, yielding significantly better performance on tasks requiring visual aesthetic judgment, spatial reasoning, and design generation.

DeepSeek essentially takes the first approach — attaching visual input capability to a powerful text reasoning model rather than redesigning the entire training pipeline around vision. That's why it struggles with design tasks that require visual aesthetic judgment.

Hallucinations: DeepSeek's Weakest Link

Another critical weakness is hallucination rate — DeepSeek's worst-performing dimension. Hallucination refers to the phenomenon where an LLM generates content that appears fluent and plausible but is actually inaccurate or entirely fabricated. Hallucinations fall into two categories: intrinsic hallucinations (output that contradicts the input information) and extrinsic hallucinations (output containing information that cannot be verified from input or training data). In agent scenarios, hallucination is especially dangerous because a model's erroneous judgment can propagate through downstream tool calls and decision chains, causing cascading failures.

The specific numbers:

• DeepSeek's hallucination-related score is only ~9% (higher is better)
• GLM 5.3 Flash scores a whopping 72%, ranking fourth — even outperforming its own larger models

In scenarios requiring factual accuracy, the GLM series is clearly more reliable, and it's also the stronger design model. This massive gap may relate to training data quality and scale — the GLM series may have received far more investment in Chinese factual data training, while DeepSeek V4 Flash, as a speed-first model, may have sacrificed some factual alignment training to maintain inference efficiency.

Selection Guide: How to Choose Between Speed and Intelligence

The final choice comes down to a classic tradeoff: do you want smarter but slower, or less smart but faster?

The tester's recommendations are highly practical:

• If you need to run multiple agents simultaneously while maintaining good speed, DeepSeek scales slightly less well than GLM 5.3 Flash but delivers roughly twice the single-model speed (DeepSeek ~80 tokens/sec vs. GLM ~30–40 tokens/sec).
• If you prioritize design capability and low hallucination rates, GLM 5.3 Flash is the safer choice.
• In high-concurrency scaling scenarios, an interesting reversal emerged in testing — while GLM led in some tests (C3, dual-stream), DeepSeek was able to "catch up and become the better model" at higher concurrency levels like C4, C5, and C6. These C3–C6 designations refer to the number of concurrent requests being processed simultaneously, reflecting throughput when multiple users or agents are calling the model at once.

It's also worth noting that the scalability results prove you don't necessarily need a 4-Spark system. Testing showed that dual Spark running TP2 performs very close to 4-Spark in single-stream scenarios — the gap only opens up under higher concurrency. That's good news for budget-conscious local deployers: two GPUs can deliver a near-four-GPU experience for single requests, and additional GPU investment only becomes necessary when you need to serve multiple agents or users simultaneously.

A Promising Experimental Release

The tester summed up his overall impression as "a coin flip" — half impressed, half disappointed. He noted plainly: "I think this model deserves a higher ranking," which stands in contrast to some benchmark scores.

DeepSeek V4 Flash Vision is a solid multimodal adaptation: it preserves the original model's speed and agentic capabilities while adding genuine image comprehension. Though design ability and hallucination rates remain clear weak points, for users running it as a daily primary model, it's "already fast enough and smart enough." With the rumored V4.5 on the horizon, DeepSeek's competitiveness in the local LLM arena is worth continued attention. This model also offers a valuable reference case for the industry: in the race to build multimodal capability, how vision is integrated matters just as much as how much vision is integrated — and the technical approach chosen will ultimately determine the ceiling of what the model can achieve.