GLM 5.3 Flash vs DeepSeek V4 Flash: Benchmarks, Hardware, and Price (2026)

The two cheapest open frontier-adjacent models of 2026 came out four weeks apart. Here’s what actually separates them.

DeepSeek V4 Flash shipped on July 31 with MIT weights and a checkpoint small enough to fit on two GPUs. GLM 5.3 Flash followed on August 26, also MIT, with image and video input and an API price that undercut DeepSeek’s. Both are sparse MoEs with a small active count, both are positioned as the volume tier under a pricier flagship, and both have people asking the same question: which one do I actually run?

The launch tables won’t answer that, because the two labs measured different things. But there’s enough shared evidence now to make a real call. Here’s the comparison.

The two models side by side

On DeepSeek’s parameter count: the model card says 284B. The 304B figure on some listings counts the DSpark speculative-decoding module that ships attached to the checkpoint. Same model.

The one benchmark they share

Z.ai’s launch table for GLM 5.3 Flash reports DeepSWE v1.1, AutomationBench, and its own code bench, compared against GLM 5.2 and Claude. DeepSeek’s table for V4 Flash uses DeepSeek’s own suite. There is no vendor-published benchmark where both models appear, so any article ranking them on a launch-table number is comparing press releases.

What does exist is a third-party ruler. Artificial Analysis runs both models through the same Intelligence Index, and as of early September GLM 5.3 Flash scores 57 to DeepSeek V4 Flash’s 52. Five points is a meaningful spread on that index, and it lines up with Z.ai’s own framing: Flash was built to sit close to its flagship on agentic and coding work rather than to be the cheap afterthought. If capability per token is the whole decision, GLM 5.3 Flash is ahead on the only shared score that exists today.

Two caveats. DeepSeek V4 Flash has been out a month longer, so its number has had more time to settle. And a composite index is not your workload; five points on an aggregate can vanish or double on a specific task, which is why the recommendation at the bottom of this post is still a bake-off.

Speed, and why DeepSeek wins it

Capability is one axis. Throughput is the other, and here the order flips. On the vendors’ own APIs, Artificial Analysis measures DeepSeek V4 Flash at around 108 output tokens per second against GLM 5.3 Flash’s 43, with time to first token close to even. Part of that is the smaller active count, 13B versus 18B. Part of it is DSpark, the speculative-decoding module DeepSeek ships with the checkpoint, which drafts up to seven tokens per step and works out of the box in both vLLM and SGLang.

Those are hosted-API measurements, so they reflect each lab’s serving stack as much as the model, and self-hosted numbers will differ. But a 2.5x spread is wide enough to matter for latency-sensitive work, and it’s the strongest argument for DeepSeek here.

GLM’s architectural edge shows up elsewhere: a hybrid sparse-plus-linear attention design that Z.ai says cuts KV cache roughly 4x versus the flagship GLM 5.3. That’s a long-context memory win, not a decode-speed win.

Hardware: two GPUs versus eight

This is the difference that decides most self-hosting conversations before benchmarks come up.

DeepSeek V4 Flash ships pre-compressed. The 166.9 GB checkpoint fits a 2x H200 pod with moderate context, and an 8-GPU node buys you the full window and QPS headroom. GLM 5.3 Flash’s FP8 checkpoint is around 306 GiB, more than two H200s hold before a byte of KV cache, so there is no two-GPU entry point. The floor is an 8-GPU Hopper-or-newer node, and the vision tower adds preprocessing overhead on top.

Quantization doesn’t close it. DeepSeek’s easy compression is already spent in the shipped format, and a flat 4-bit GLM is still roughly 160 GB before overhead. Both stay multi-GPU no matter what you do; DeepSeek just starts at a smaller multi. If you’re renting, that’s a two-GPU pod versus a full node, every hour the model is up.

Price: GLM’s API is cheaper, DeepSeek’s hosting is cheaper

The API math favors GLM. Z.ai launched Flash at $0.15 per million input tokens and $0.50 output, with cached input at $0.03. DeepSeek raised its rates on August 17 to $0.44 in and $1.32 out at peak, halved off-peak, with a steep cache discount. On Artificial Analysis’s blended price, that’s roughly $0.10 per million for GLM against $0.23 for DeepSeek, and GLM lands on the intelligence-per-cost frontier as a result.

The self-hosting math favors DeepSeek, for the hardware reason above. A 2x H200 pod running V4 Flash flat-rate can beat either API on sustained volume; an 8-GPU GLM node has to clear a much higher bar against a $0.50 output price. In practice that pushes GLM 5.3 Flash toward the API and leaves DeepSeek V4 Flash as the only one of the two where self-hosting is a plausible default.

Modalities and the rest

GLM 5.3 Flash takes image and video input natively, the first model in the GLM-5 line that does. DeepSeek V4 Flash is text only. If vision is in the workload at all, the comparison ends here.

Both are MIT, so fine-tuning and commercial use are clean on either. Both landed with day-one vLLM and SGLang support. DeepSeek’s serving path has an extra month in the wild behind it, which counts for something when you hit the first odd engine bug; the vLLM vs SGLang comparison covers the engine choice for either.

Which one to use

Choose GLM 5.3 Flash if you’re API-first and want the higher shared score at the lower price, if you need image or video input with open weights, or if long-context agent work is the core of the workload. The 8-GPU floor only matters if you insist on hosting it yourself.

Choose DeepSeek V4 Flash if you plan to self-host, if decode speed is a hard requirement, or if you want the option with the longest independent track record. Its checkpoint is the most approachable of any open model at this capability level, and that’s a structural advantage GLM can’t match.

If you’re running both patterns, which most production teams are, the split writes itself: GLM 5.3 Flash through an API for the capability-sensitive traffic, DeepSeek V4 Flash self-hosted for the flat-cost volume tier. Since both speak OpenAI-compatible interfaces, that’s routing configuration, not two integrations.

How to run each today

DeepSeek V4 Flash: self-host on a 2x H200 pod using our deployment tutorial, or call it through Yotta AI Gateway, where V4 Flash and V4 Pro are both live with flat pricing and no peak-hour windows.

GLM 5.3 Flash: Z.ai’s API at launch pricing, or self-hosted on an 8-GPU node using the same serving stack our GLM 5.2 SGLang guide walks through, with multi-GPU capacity by the hour. The flagship GLM 5.3 is live on the Gateway alongside GLM 5.2, so a three-way test against DeepSeek is a model-string change.

Frequently asked questions

Is GLM 5.3 Flash better than DeepSeek V4 Flash? On the only shared third-party score, Artificial Analysis’s Intelligence Index, GLM 5.3 Flash leads 57 to 52 as of early September 2026. DeepSeek V4 Flash is roughly 2.5x faster on the vendors’ hosted APIs and far cheaper to self-host. Neither vendor’s launch table includes the other model.

Which is cheaper? By API, GLM 5.3 Flash: $0.15 in and $0.50 out versus DeepSeek’s $0.44 and $1.32 at peak after its August price increase. By self-hosting, DeepSeek V4 Flash, because its 166.9 GB checkpoint runs on two H200s while GLM needs an 8-GPU node.

Which is faster? DeepSeek V4 Flash, by a wide margin on hosted-API measurements: around 108 output tokens per second against 43 for GLM 5.3 Flash. Smaller active parameters and built-in speculative decoding both contribute.

Can I run either on a single GPU? No. The smaller of the two checkpoints is 166.9 GB. Two H200s is the floor for DeepSeek V4 Flash; GLM 5.3 Flash needs eight datacenter GPUs.

Which one handles images and video? GLM 5.3 Flash. DeepSeek V4 Flash is text only.

Are both open source? Both ship full weights on Hugging Face under MIT, so fine-tuning and commercial deployment are unrestricted on either.

Are they on Yotta? DeepSeek V4 Flash and V4 Pro are live on Yotta AI Gateway alongside the flagship GLM 5.3 and GLM 5.2, one API key with OpenAI-compatible endpoints. This post will note it when GLM 5.3 Flash joins the catalog.

Bottom line

GLM 5.3 Flash is the smarter model on the one score they share and the cheaper one to call. DeepSeek V4 Flash is the faster model and the only one of the two most teams can realistically host. That’s not a tie; it’s two different products that happen to share a price tier, and the right one depends on whether your bottleneck is capability, latency, or the hardware bill.

The full breakdowns: GLM 5.3 Flash hardware requirements, DeepSeek V4 Flash hardware requirements, and the three-way Flash-class comparison that adds Qwen 3.8-Flash-Next. When you’re ready to measure instead of read, DeepSeek V4 Flash is live on Yotta AI Gateway, and multi-GPU pods by the hour cover the self-hosted side.