High-speed serving tier of GLM-5.3-Flash for interactive coding agents and multimodal tool loops at up to 200 tokens/s.
Modality:
Text
Image
PDF
model ID
z-ai/glm-5.3-flashx
Output Speed *
65.00
tok/s
Intelligence Index *
44.8
/ 100
Context Window *
1048576
tokens
Input price
2.22
Anytoken
Output price
7.5
Anytoken
GLM 5.3 FlashX: Fast Multimodal Inference for Coding Agents and Tool Loops
GLM 5.3 FlashX is Z.ai's high-speed serving variant of GLM-5.3-Flash, the first native multimodal model in the GLM-5 series. It runs the same 320B-total / 18B-active Mixture-of-Experts weights as Flash but is optimized at the inference layer for throughput, reaching up to 200 tokens/s. It sits in the budget, speed-focused band of the GLM-5.3 lineup rather than the flagship reasoning tier. The combination of a 1M-token context window, image and video input, always-on reasoning, and fast generation makes it best suited to interactive coding agents and long-horizon tool loops where latency matters.
Start building with GLM 5.3 FlashX through the AnyAPI.ai API.
Performance
Why FlashX Trades Nothing on Weights but Everything on Speed
GLM 5.3 FlashX is built for interactive latency: Z.ai reports peak generation speeds of up to 200 tokens per second, which independent gateways cite as roughly 5x faster inference than standard GLM-5.3-Flash. Because FlashX serves the identical 320B/18B MoE weights as Flash, it inherits Flash's near-frontier coding profile rather than defining a new one. For developers, this means the perceived responsiveness of agents, code streaming, and multi-step tool loops improves substantially without changing the underlying model's reasoning behavior. The practical consequence: FlashX is a serving choice, not a capability upgrade.
Benchmarks
What the Underlying Flash Weights Score Independently
FlashX shares weights with GLM-5.3-Flash, so Flash's independent benchmarks are the relevant reference. On Artificial Analysis, GLM-5.3-Flash records a Coding Index of 71.5 and an Agentic Index of 50.9. Z.ai's own evaluation reports 84.3 on Terminal-Bench 2.1 and 63.4 Pass@1 on DeepSWE v1.1, up sharply from GLM-5.2's 46.2. Independent reviewers place its Terminal-Bench 2.1 result near Claude Opus 4.8 (85.0). These figures describe the Flash weights; FlashX's documented difference is serving speed, so coding or reasoning quality should not be assumed higher than Flash.
Output Speed
*
65.00
tok/s
Intelligence Index
*
44.8
/ 100
MMLU *
Broad world knowledge and problem-solving
0
%
GPQA *
PhD-level scientific reasoning across physics, biology, chemistry.
92
%
HLE *
Adherence to multi-step structured instructions.
42
%
LiveCodeBench *
Tool-calling reliability in long agentic loops.
0
%
Technical Specifications
What the model supports
GLM 5.3 FlashX accepts text, image, video, and file input and returns text only. It supports a 1M-token (1,048,576) context window with a maximum output of 128K tokens — two distinct limits that should not be conflated when sizing long-document or long-generation workloads. Reasoning is always on, with effort controllable but not fully disabled. The hybrid sparse-plus-linear attention architecture is what makes 1M-token inference economical, but the model cannot emit images, audio, or video. Function calling and structured JSON output are supported, and the API follows a Chat Completions-compatible format.
Verified Specifications —
GLM 5.3 FlashX
*
Input modalities
Text
Image
PDF
output modalities
Text
Image
Context window
1048576
tokens
Maximum output tokens
131072
Reasoning
Yes
Knowledge cutoff
September 2026
Pricing (standard)
2.22
AnyTokens in
/
7.5
AnyTokens out
Comparison
GLM 5.3 FlashX vs GLM 5.3 Flash: What Actually Changes?
GLM 5.3 FlashX and GLM 5.3 Flash are the same model family with the same weights — a 320B-total / 18B-active MoE, 1M-token context, and native multimodal input. The only documented difference is serving: FlashX is Z.ai's high-speed inference variant, reaching up to 200 tokens/s against Flash's lower throughput. Because the weights are identical, coding, reasoning, and agentic quality are effectively the same. The practical decision is therefore not about intelligence but about whether your workload needs faster streamed responses and is willing to accept FlashX's different pricing tier.
Dimension
GLM 5.3 FlashX
GLM-5.3-Flash
Context window *
1048576
tokens
1048576
tokens
Output speed *
65.00
tok/s
52.01
tok/s
Intelligence Index *
44.8
41.8
Input pricing
2.22
AnyToken
0.45
AnyToken
Output pricing
7.5
AnyToken
1.5
AnyToken
Knowledge cutoff *
September 2026
August 2026
Choose GLM 5.3 FlashX when response latency is user-visible — interactive coding agents, live tool loops, and real-time applications where throughput drives the experience. Choose GLM 5.3 Flash when you are running batch or background jobs, cost-sensitive high-volume generation, or workloads where a few hundred extra milliseconds do not matter, since Flash occupies the lower-cost serving tier. Neither choice changes answer quality; select based on latency sensitivity and economics.
Limitations & Trade-offs
1
Speed-only variant, not a capability upgrade. FlashX uses the identical weights as GLM-5.3-Flash; Z.ai positions it as an inference and infrastructure optimization delivering up to 200 tokens/s. No published evidence shows higher coding or reasoning quality than Flash. Teams expecting a smarter model will be disappointed — if you need more raw capability, the flagship GLM-5.3 reasoning tier is the appropriate upgrade, not FlashX.
2
Reasoning cannot be fully disabled. Reasoning is always on for FlashX, and lowering reasoning_effort does not turn it off. This adds token overhead and latency variability on simple requests where you may not want any thinking budget. For high-volume, low-complexity classification or extraction tasks, a model that lets you switch reasoning off entirely may be cheaper and more predictable.
3
Text-only output despite rich input. FlashX accepts text, image, video, and file input but generates only text. It cannot produce images, audio, or Office files directly — document or media artifacts come from agent tools, not the model. Applications needing native image or audio generation must pair FlashX with separate generation models.
4
Higher per-token price than Flash. FlashX occupies a more expensive serving tier than standard GLM-5.3-Flash. For long-output or high-volume background generation where latency is not user-facing, the speed premium adds cost without delivering a quality benefit, making Flash the more economical default for those jobs.
Best-Fit Workloads
Where this model earns its place
01
Interactive coding agents
FlashX's up-to-200 tokens/s throughput makes streamed code generation feel responsive inside IDE assistants and autonomous coding agents. The underlying Flash weights score 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE v1.1, giving near-frontier coding behavior. Native visual coding — observing rendered interfaces and feedback — supports frontend and UI iteration loops where fast turnaround compounds across many steps.
02
Long-horizon tool loops
Agentic workflows that call the model repeatedly for planning, tool invocation, and review benefit directly from higher throughput, since latency accumulates across every step. With multi-step function calling and a 1M-token context, FlashX can sustain long-running agent sessions that coordinate code, browsers, and documents without losing earlier context. Faster inference reduces end-to-end wall-clock time for these loops.03
Real-time multimodal assistants
Applications that accept screenshots, images, or documents and must respond quickly — such as visual debugging helpers or interactive document Q&A — fit FlashX's combination of image/video/file input and fast text generation. The 1M-token window lets full documents or interface context sit in a single request. Note that output is text only, so any generated files must come from downstream tools.04
Large-codebase and long-document RAG
The 1M-token context window, made economical by hybrid sparse-plus-linear attention, lets FlashX ingest entire repositories or lengthy document sets without chunking. Combined with fast generation, this suits retrieval-augmented workflows where latency to first useful answer matters. Keep the 128K maximum output limit in mind: large inputs do not imply proportionally large generations.Pricing in anytokens via AnyAPI
Input
2.22
₳
Output
7.5
₳
Cache write
—
₳
Cache read
0.54
₳
Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04