Verdict: The best open-source LLM for coding in 2026 is still Qwen3-Coder-480B-A35B-Instruct for pure patch generation under an Apache-2.0 licence, with GLM-5.2 ahead on long agentic runs and DeepSeek V4 Flash the clear cost leader. The real change since our full head-to-head on Qwen3-Coder and GLM-5.2: DeepSeek shipped a vision-capable Flash model (deepseek-v4-flash-vision-exp), so the cheapest serious open coding stack can now also see screenshots, error dialogs, and UI states.
TL;DR - Last verified: 2026-08-23. Prices are volatile and were re-checked against the official rate card today.
- Default pick: Qwen3-Coder-480B (Apache-2.0, 69.6% on SWE-bench Verified)
- Agentic pick: GLM-5.2 (MIT licence, leads on long-horizon agent benchmarks)
- Cost pick: DeepSeek V4 Flash - from $0.22/1M input in off-peak hours, MIT-licensed weights
- New:
deepseek-v4-flash-vision-expadds native image input at the same price as text Flash
Which open-source coding model should you actually pick?
Pick by your constraint, not by a leaderboard. The three serious options diverge on exactly three axes, and each wins one outright.
| Qwen3-Coder-480B | GLM-5.2 | DeepSeek V4 Flash | |
|---|---|---|---|
| Licence | Apache-2.0 | MIT | MIT |
| Size | 480B total, 35B active (MoE) | Large MoE, open weights | 284B total, 13B active (MoE) |
| Context | 256K (1M with extrapolation) | Large context | 1M tokens, up to 384K output |
| Known strength | SWE-bench Verified 69.6% | Long-horizon agent runs | Cost per token, throughput |
| Vision input | No | No | Yes, via deepseek-v4-flash-vision-exp
|
| Best when | You need the cleanest enterprise licence | Your agent runs for hours across a repo | Token budget is the binding constraint |
Sources: Qwen's release notes and licence at Qwen3-Coder-480B-A35B-Instruct on Hugging Face; DeepSeek's model family on the official Models & Pricing page; GLM-5.2's MIT licence and agentic benchmark position are covered in our existing Qwen3-Coder vs GLM-5.2 comparison with primary links.
What did DeepSeek actually add with vision?
DeepSeek introduced a vision-capable variant of its V4 Flash model, exposed in the API as deepseek-v4-flash-vision-exp. Per the official vision guide (checked 2026-08-23), it accepts images alongside text over the same OpenAI-compatible Chat Completions API, the Anthropic-compatible /anthropic route, and the Responses API.
Three ways to send an image:
-
Base64 inline as a
data:URL inside animage_urlblock. Counts toward the 48 MiB request body limit. - A public URL the model downloads itself - URL up to 8,192 characters, image up to 32 MiB.
-
A Files API
file_idfor reuse or large images - up to 64 MiB per image, exempt from the 32 MiB URL/base64 check.
Practical details that matter for a coding assistant:
- Images are billed as input tokens, capped at 384 tokens per image regardless of original resolution (a 2000px and a 5000px image cost the same), per the same vision guide.
- Formats are JPEG, PNG, GIF, and WebP, detected from file content.
- Images only go in
usermessages; up to 600 images per request. - None of the other DeepSeek models accept images - you get a 400 if you try.
The exp suffix is doing honest work: this is DeepSeek's experimental vision release, so quality and behaviour can shift before it graduates to a stable alias.
How much does DeepSeek V4 Flash cost now?
From the official DeepSeek Models & Pricing page, checked 2026-08-23. All figures per 1M tokens, and identical for the vision model - images are just converted into input tokens:
| Off-peak | Peak | |
|---|---|---|
| Input, cache hit | $0.007 | $0.014 |
| Input, cache miss | $0.22 | $0.44 |
| Output | $0.66 | $1.32 |
Off-peak hours are everything outside 01:00-04:00 and 06:00-10:00 UTC. If you run an agent all night on off-peak pricing and hold its context in cache, an 8-hour coding session costs you single-digit cents in tokens - that is the whole pitch, and it is why Flash showed up as the default sub-agent model in so many agent stacks, including the plugin-driven DeepSeek Harness framework.
For comparison, Qwen3-Coder on hosted providers runs roughly $0.22-$0.30/1M input and $1.00-$1.80/1M output depending on the provider (per OpenRouter's Qwen3-Coder listing, checked 2026-08-23). DeepSeek still wins that matchup on output price even at peak.
If you want to try Flash without paying at all, we verified five free access paths in our DeepSeek V4 Flash free setup guide.
Why does vision matter for a coding model?
Because a working agent keeps bumping into things that are pictures. The screenshots an agentic coder has to interpret - a traceback rendered in a terminal, a broken layout, a CI badge, an error dialog - are often faster to hand over as an image than to transcribe into text.
Three concrete workflows that change when the cheap sub-agent model gains vision:
- Debug from screenshots. An orchestrator captures the failing UI and hands the screenshot straight to the sub-agent. No OCR step, no separate vision call, no context loss between "what the screen shows" and "what the code says."
- Verify generated front-ends. The agent renders its own output, takes a screenshot, and self-corrects: missing padding, clipped text, a flag drawn in the wrong place. Vision closes the loop without a human in it.
- Read diagrams and scans. Architecture sketches, PDF pages, receipts, whiteboard photos - a coding model that sees them can fold them into the same reasoning as the code, at 384 tokens each.
Vision does not make the model smarter at writing code. It removes a transcription bottleneck between the agent and its environment, at roughly the price of a short paragraph of extra input.
Is the vision model good enough to be your daily driver?
Honest answer: good enough to try, not yet good enough to default to. Our read, from the published specs and early behaviour:
- Speed holds up. Flash is the throughput-optimised sibling in the family (284B total parameters, only 13B active per token), and vision input rides the same fast path. It feels like a sub-agent model, not a flagship.
-
Complex generated artifacts are hit or miss. Long, playable game builds from a single prompt still come out janky - a reminder that
explabels exist for a reason. Treat single-shot app generation as a demo trick, not a workflow. - The right architecture is orchestrator plus sub-agents. Use a stronger model (V4 Pro, or the GLM/Qwen picks above) as the planner, and V4 Flash Vision as the fast, cheap worker that handles screenshots, classification, and routine edits. That split is where the pricing shines.
What this means for you
If you run coding agents, add vision to your cheapest tier. Specifically:
- Point your sub-agent calls at
deepseek-v4-flash-vision-expfor any step that consumes screenshots, rendered output, or scans - the input price is unchanged from text Flash. - Send images by URL or Files API where you can; base64 inflates the 48 MiB body limit fast.
- Keep Qwen3-Coder as your licence-safe patch generator and GLM-5.2 as your long-run agent. Nothing about DeepSeek's vision release changes that ordering - it widens where the budget option fits.
FAQ
Q: Does DeepSeek V4 Flash support image input?
A: Yes, through the deepseek-v4-flash-vision-exp model ID. It accepts JPEG, PNG, GIF, and WebP images via base64, a public URL, or the Files API, over both the OpenAI-compatible and Anthropic-compatible endpoints. Base Flash and V4 Pro return a 400 error if you send them images.
Q: How much do images cost on DeepSeek's vision model?
A: Images are converted to input tokens and billed at the standard input rate, capped at 384 tokens per image regardless of resolution. At off-peak cache-miss rates ($0.22 per 1M input tokens), one image costs a fraction of a cent.
Q: Is DeepSeek V4 Flash open source?
A: The weights are published under the MIT licence, free for commercial use. That puts it in the same permissive tier as GLM-5.2 and alongside Apache-2.0 Qwen3-Coder.
Q: Which open-source model is best for coding in 2026?
A: There is no single winner. Qwen3-Coder-480B leads on SWE-bench Verified (69.6%) with the cleanest Apache-2.0 licence; GLM-5.2 leads on long agent runs; DeepSeek V4 Flash is the cheapest by a wide margin and now the only one of the three with native vision. Pick by your constraint.
Q: Should DeepSeek V4 Flash Vision be my main coding model?
A: No - the exp label is accurate. Use it as the fast, cheap sub-agent for screenshots, classification, and routine edits, with a stronger model (V4 Pro, GLM-5.2, or Qwen3-Coder) as the orchestrator doing the hard planning.
Q: What is the context window of DeepSeek V4 Flash?
A: One million tokens of context with up to 384K tokens of output, for both the standard and vision variants, per DeepSeek's Models & Pricing page.
Sources
- DeepSeek API Docs - Models & Pricing (checked 2026-08-23): model IDs, prices, context, peak/off-peak schedule
-
DeepSeek API Docs - Vision (checked 2026-08-23):
deepseek-v4-flash-vision-exp, input methods, image token cap, limits - Qwen3-Coder-480B-A35B-Instruct on Hugging Face: weights and licence
- OpenRouter - Qwen3 Coder (checked 2026-08-23): hosted provider pricing
Updates & Corrections
- 2026-08-23 - Initial publication. DeepSeek pricing re-verified against the official rate card; vision guide limits re-verified.
Researched and drafted with AI agents; reviewed and fact-checked under human editorial oversight at The Tech Archive.
Top comments (0)