Choosing a Multimodal Chinese Model: Vision and Video

Short answer. Multimodality has migrated down the price stack in the 2026 lineup: the cheapest tiers are now among the most capable. Ten models record vision input, five record video input, and four record computer use — and notably, DeepSeek’s $0.66 flagship V4-Pro and Zhipu’s $1.40 GLM-5.3 are text-only while their $0.15 flash siblings (V4.1-Flash, GLM-5.3-Flash) see images, and GLM-5.3-Flash adds video and computer use. No tracked model records an audio-input flag. The practical rule: pick the capability first, then notice that capability is per-model, not per-price-tier.

Decision criteria

Criteria	Relevance / Notes
Vision (image input)	10 models: V4.1-Flash, Doubao Pro/Turbo/Evolving, GLM-5.3-Flash, Kimi K2.6, Kimi K3, MiniMax M3, Qwen3.8-Flash, Qwen3.8-Max
Video input	5 models: GLM-5.3-Flash, Kimi K3, MiniMax M3, Qwen3.8-Flash, Qwen3.8-Max — the rarest mainstream flag
Computer use (GUI)	4 models: Doubao Pro/Turbo/Evolving, GLM-5.3-Flash — every computer-use model also has vision
Audio input	No tracked model records an audio flag
Flagship/vision inversion	V4-Pro and GLM-5.3 are text-only; their $0.15 flash siblings carry vision
Price of vision	$0.15 input (V4.1-Flash, GLM-5.3-Flash) — no reason to pay flagship prices for image understanding

The modality matrix

Model	Vision	Video	Computer use	Price (input/output)
DeepSeek-V4.1-Flash	✓	—	—	$0.15 / $0.60
DeepSeek-V4-Pro	—	—	—	$0.66 / $1.98
GLM-5.3	—	—	—	$1.40 / $4.40
GLM-5.3-Flash	✓	✓	✓	$0.15 / $0.50
Kimi K3	✓	✓	—	$3.00 / $15.00
MiniMax-M3	✓	✓	—	$0.30 / $1.20
Qwen3.8-Max	✓	✓	—	$2.00 / $6.00
Qwen3.8-Flash	✓	✓	—	$0.15 / $0.47
Doubao Seed 2.1 Pro	✓	—	✓	¥6 / ¥30 (CNY)
Doubao Seed 2.1 Turbo	✓	—	✓	¥3 / ¥15 (CNY)
Doubao Seed Evolving	✓	—	✓	¥6 / ¥30 (CNY)
Kimi K2.6	✓	—	—	$0.95 / $4.00

Entity routing

Route by the modality you actually need.

Scenario	Best-documented fit	Why
Vision at minimum cost	DeepSeek-V4.1-Flash or GLM-5.3-Flash	$0.15 input with vision
Video input	Qwen3.8-Max / Qwen3.8-Flash / Kimi K3 / MiniMax-M3 / GLM-5.3-Flash	The five video-capable models
Computer use (model-level)	Doubao Seed family or GLM-5.3-Flash	Ark (CNY, cn-beijing) or Z.ai (USD, open)
Vision + video + 1M output	Kimi K3	1M/1M, $3/$15
Full perception stack, closed platform	Doubao Seed 2.1 Pro	Vision + computer use + agent capability
Open-weight multimodal	GLM-5.3-Flash or MiniMax-M3	Apache-2.0 / community license
Video understanding benchmarked	Kimi K3	Video-MME (with subtitles) 90.0

What the evidence shows

The capability map reflects two product theories. Vendors that ship open weights (Zhipu, Moonshot, MiniMax) treat multimodality as a spec to publish per model; vendors that sell closed APIs treat capabilities as product releases — ByteDance’s Seed family updates capability through version snapshots (260628 → 260915) rather than new model cards. The flag set is therefore not a pure capability map but a capability map filtered through disclosure policy.

The flagship/vision inversion falsifies the old “pay more, get more modalities” assumption at two of six vendors: DeepSeek’s $0.66 flagship V4-Pro is text-only while its $0.15 V4.1-Flash sees images; Zhipu’s $1.40 GLM-5.3 is text-only while GLM-5.3-Flash carries vision, video and computer use. In both cases the reasoning-specialized flagship predates or deliberately omits the multimodal stack, and the newer cheap model inherits the full capability set.

Video input is the rarest mainstream flag (five models), and Qwen is the only vendor where both flash and flagship carry video. Computer use is a ByteDance stronghold — three of four computer-use models are Doubao Seed — though the distinction between a model-level computer-use flag and an agent product that orchestrates screen control with a text model plus external tooling matters, and the agents collection tracks that separately. Every computer-use model also has vision and agent capability — computer use sits on top of a full perception stack, never standalone.

China AI Hub analysis indicates the trajectory is a one-way migration of multimodal stack from premium to entry-level: DeepSeek’s first vision API model arrived in the flash tier (V4.1-Flash, 2026-09-10), and Zhipu moved vision + video + computer use from premium to entry-level in a single generation (GLM-5.2 text-only in June → GLM-5.3-Flash multimodal in August). Vision is now baseline, so it ships where volume is.

Selection procedure

Work through these steps in order.

Identify the exact modality. Vision alone is broad and cheap; video input is rare (five models); computer use is rarest (four models) and concentrated in ByteDance plus GLM-5.3-Flash. Name the capability you need before shortlisting, because the model sets barely overlap.
Do not assume flagship means multimodal. At DeepSeek and Zhipu the reverse is true: V4-Pro and GLM-5.3 are text-only while their $0.15 flash siblings carry vision. Check the per-model flag, not the price tier.
If you need vision at minimum cost, the field is two models. DeepSeek-V4.1-Flash and GLM-5.3-Flash are both $0.15 input with vision; GLM adds video and computer use at the same price.
If you need video input, weigh price against output headroom. Qwen3.8-Flash is the cheapest video-capable API ($0.15/$0.47); Kimi K3 combines video with 1M output and posts Video-MME 90.0.
If you need computer use, it is a platform decision, not just a model one. The Doubao Seed family (Ark, CNY, cn-beijing) versus GLM-5.3-Flash (Z.ai, USD, open weights) are not substitutes — they live on different platforms with different licensing.
Treat a missing flag as undocumented, not absent. A model without a vision flag may still accept images in some product surface; verify against the model page before concluding a capability is missing.

Limitations

Flag absence is not capability absence: a model without a vision flag may still accept image input in some product surfaces, and the database records what official pages state. Computer-use flags are model-level declarations; agent products can deliver screen control without a model-level flag, so the flag map undercounts the practical agent landscape. Flag granularity varies — some vendors enumerate modalities in free text rather than structured lists, so the database normalizes conservatively and may lag a vendor’s silent capability additions. No model records an audio-input flag, but audio output and separate audio models are outside this collection’s scope. Video input and video generation are different capabilities and are tracked separately.

Sources

Labels used above: Official fact (capability flags and pricing from primary sources), Vendor-reported claim (modality support as published by the vendor), and China AI Hub analysis (our synthesis, introduced as such). No third-party evaluation evidence is currently recorded for multimodal capability in this collection.