DeepSeek bolted a vision layer onto its fast model without touching anything else, and the surprise landed where nobody was looking: the text scores went up too. This V4-Flash-Vision test pulls together the published evaluations, the real billing terms and what the variant actually changes inside an agent loop.
Key Takeaways
- The vision variant beats the text-only model on six text benchmarks out of seven, which was never the stated goal.
- It clears Opus 4.8 on three agentic lines and stays twelve points behind on repository-scale code.
- Every image counts as 384 tokens at V4-Flash rates, with a ceiling of 600 images per request.
Have an AI Sum Up This Article
ChatGPTA vision graft that lifted the text scores
The first finding of this V4-Flash-Vision test concerns what actually shipped. This is not a new model, it is an extension: the text base gains the ability to read images, and the rest of its behaviour is meant to carry over untouched.
The model is served under the identifier deepseek-v4-flash-vision-exp. Images arrive as base64, as an external URL or through the Files API, in JPEG, PNG, GIF and WebP, which covers most of what a document pipeline will throw at it.
The integration choice says something about the strategy. Rather than standing up a separate multimodal family with its own release cadence and its own price sheet, the lab pushes vision into a tier customers have already wired in, where the integrations exist and the cost per call is known.
The starting point matters as much as the graft. This is V4 Flash, the model that landed a point behind GPT-5.6 Luna at 60% lower cost, so a base tuned for price-performance rather than for the summit.
Where this test gets genuinely interesting is that adding vision appears to have cost the text side nothing at all. Across seven text benchmarks, the multimodal variant beats the text-only model on six of them.
Two gaps stand out. Toolathlon-Verified climbs 5.6 points and DeepSWE 4.9 points against the text-only release. Multimodal training that improves tool use is not the result the product page leads with.
The sharpest jump shows up against its own predecessor. On ApexBench at Pass@1, the vision variant reaches 36.5 where V4-Flash 0731 topped out at 26.2, which puts it three points off Opus 4.8 on that specific line.
One caveat rides on all of these numbers, and it is written into the model name. The Exp suffix marks an experimental milestone, with results measured in the in-house harness in minimal mode and no independent verification available at this stage.
We therefore read those tables for what they are: a quantified statement of intent from the vendor, useful for framing a test and nowhere near enough to sign off a migration.
What vision actually unlocks inside an agent loop
For a team running agents in production, the leaderboard is not what decides adoption. What decides it is handling screenshots, scanned PDFs and charts inside the same loop as the text, without routing out to a second and pricier model.
That is exactly what the graft approach buys. Migrating from V4-Flash means no tool rewrite and no system prompt recalibration, since the agentic behaviour of the base model carries over as is.
On agent ground, the three wins the variant posts against Opus 4.8 across eleven benchmarks land in the right place. Agents’ Last Exam reads 27.3 against 25.7, ZeroBench at Pass@5 reads 35.0 against 34.0, and DeepSWE clears by 1.3 points.
The margins are thin, and that thinness is precisely the message worth taking away. A model at this price point reaching a frontier model on multimodal agent tasks, even by a point and a half, moves the conversation from capability to cost.
The practical constraint comes down to two parameters. Each image is capped at 384 tokens, and a single request accepts up to 600 images, which bounds the cost of a document batch in advance instead of letting it drift with file resolution.
For a team costing out a pipeline, that predictability is worth more than a benchmark point. It turns a variable invoice into a budget line you can calculate before anything reaches production.
The 600-image ceiling opens a use case this V4-Flash-Vision test rates highly. A full batch of invoices, layouts or captures goes out in one call, with no application-side chunking and no sub-request orchestration to maintain.
The cap cuts the other way too. A dense document compressed into 384 tokens loses detail, and tasks that require reading a crowded table closely will suffer more than standard screenshot work.
Before switching anything, the number to measure is your own. Run a representative sample of your documents through both the current routing and the vision variant, then compare extraction accuracy rather than benchmark position, because the 384-token compression hits some file types far harder than others.
Reports from teams already running the V4 Flash family in production add a nuance that matters. The models are described there as partly benchmark maxed, strong on standard evaluations and less consistent in everyday use.
That reservation matches the read we took when the first tables came out. The ranking flatters more than the usage average does, and a week of real workloads almost always corrects the impression left by one well-placed benchmark line.
More articles on Horizon
- Ox Alpha: A Free AI Model Arrives With No Vendor
- DeepSeek V4-Flash-Vision Closes In on Opus 4.8
- Claude Security Now Scans Code With Mythos 5
Our verdict, workload by workload
Where the model drops off, the gap is not symbolic. On NL2Repo, which measures repository-scale code generation, it posts 57.7 against 69.7 for Opus 4.8, twelve points behind on the exercise closest to a developer’s actual day.
On DSBench-Hard the deficit runs to around eight points. Elsewhere the distance narrows to the point of irrelevance: Terminal Bench 2.1 puts it at 83.9 against 85.0, and Chartography at 64.3 against 65.0.
One methodological detail is worth holding on to before reading those gaps as settled. The measurements come out of the vendor’s harness in minimal mode, meaning the conditions most favourable to a clean run, far from the improvised prompts and cluttered environments of daily use.
The spread draws a readable map. The model holds the comparison on short, tool-driven, multimodal tasks, and loses it the moment the work demands holding an entire architecture in mind across dozens of files.
Our verdict for a document support agent is firmly positive, and it does not depend on the overall ranking. Reading captures, extracting from scanned PDFs, sorting charts, chaining tool calls: on that profile the capability-to-price ratio is hard to beat today.
Our verdict for repository-scale code generation is the opposite. Twelve points behind on NL2Repo is not recovered by a lower rate, because the human correction time eats whatever the token savings delivered.
Between the two sits a case for waiting. A team already settled on V4-Flash for text has every reason to switch, since the migration costs nothing and six text benchmarks out of seven improve, but it should keep a fallback model for the long tasks.
Competitively, this V4-Flash-Vision test confirms where the pressure actually lands. Not on peak capability, which frontier models keep, but on vendors billing a vision component on top of a text model, now undercut by a single brick charged at one tier’s rate.
The invoice, though, is not read off this one catalogue line. DeepSeek tunes its margins tier by tier, and the move of V4-Pro to peak and off-peak billing pushed some cache lines up by more than 1,100%.
There is also a timing question worth planning around. An experimental release moves fast, and the version measured today may not be the version billed in six weeks, which argues for a short evaluation now and a firm decision once a stable identifier replaces the current one.
Our closing recommendation rests on that pricing instability as much as on the measured scores. The model is worth testing this week against your own workloads, and the contractual commitment is worth holding back until the experimental suffix drops and the rates settle.
Follow the story on Horizon.
