Skip to content

Inco AI launches Day-0 support for GLM 5.3

Z.AITokenRouter.com

GLM 5.3 — Cumulative Serving Speedup

End-to-end serving throughput

Concurrency 1
Figure 1. End-to-end GLM 5.3 throughput at concurrency 1 with DFlash 2, NVFP4, and Inco Engine improvements, relative to native FP8 autoregressive decoding.

Overview

GLM 5.3 arrives today with day-0 serving support from Inco AI on TokenRouter compute. The release pairs Z.ai's new model with our core inference optimization technology:

  1. A DFlash 2 checkpoint for speculative decoding.
  2. An NVFP4 checkpoint for efficient Blackwell inference with native accuracy.
  3. Inco Engine for end-to-end inference performance.

As a day-0 launch partner, we are releasing both checkpoints alongside the model endpoint. Together, the checkpoints and Inco Engine deliver up to 4.4× throughput versus the native FP8 checkpoint with autoregressive decoding at concurrency 1.

DFlash 2 for GLM 5.3

DFlash 2 is a parallel drafter for speculative decoding: it predicts candidate tokens in one pass and lets the target model verify them as a block.

Diffusionisgood⟨mask⟩⟨mask⟩⟨mask⟩Independent Top-1 Picksfor✗same word, twiceposition 1decodingspeculativeslow⋮position 2decodingthinkingmodels⋮position 3⟨eos⟩again…⋮all adjacent pairs scored at once → one path keptaccepted outputfor
target-decoded tokenmask tokenaccepted draftselected path
Figure 2. DFlash 2 keeps each draft position's top candidates and uses a lightweight path selector to trace one coherent block for the target model to verify.

Acceptance length (AL) measures how many tokens each draft–verify cycle yields, including the verifier's next token. A longer accepted block means fewer full target-model passes for the same output. DFlash 2 improves AL by keeping the parallel draft design while selecting a more coherent path through each position's candidates.

Acceptance length is only half of the serving result: a drafter also adds work to each cycle. For GLM 5.3, we therefore report end-to-end decode throughput separately, comparing the model's native MTP path and DFlash 2 against autoregressive decoding. At concurrency 1, DFlash 2 reaches 383.3 output tok/s on MATH-500, 366.6 on GSM8K, and 363.7 on HumanEval.

GLM 5.3 — DFlash 2 performance

MTPDFlash 2
Figure 3. GLM 5.3 DFlash 2 acceptance length and concurrency-1 decode speedup versus native MTP on the three datasets with the highest DFlash 2 throughput speedup. Broken axes omit the uninformative interval after the 1.0 baseline; speedup is relative to autoregressive decoding.

NVFP4 for Blackwell

We are also releasing an NVFP4 checkpoint for efficient Blackwell deployment, with accuracy matching the native FP8 checkpoint across the evaluation suite.

PrecisionGPQA DiamondAIME 2025MATH-500HLEAA-LCR
FP891.194.395.635.973.6
NVFP491.295.195.235.273.0
Table 1. GLM 5.3 accuracy for the native FP8 and Inco NVFP4 checkpoints.

Get access

Preview

Try GLM 5.3 in the browser or connect through the Inco API.

Inco Playground

Included in preview

  • Browser playground
  • OpenAI and Anthropic API dialects
  • Streaming responses
  • 1M-token context window

List pricing

per 1M tokens
Input
$1.40
Cached input
$0.26
Output
$4.40

Get checkpoints

Download the checkpoints from Hugging Face. Serving instructions are on each model card.

DFlash 2NVFP4

Compute for GLM 5.3

Compute for this launch came from TokenRouter. We trained the DFlash 2 drafter and produced the NVFP4 checkpoint on its Blackwell clusters, and we serve the day-0 endpoint from the same pool. You can reach the endpoint with an existing TokenRouter key.

The bottom line

Ten days ago, we called DFlash 2 the first piece of an end-to-end serving stack. Launches like this are what the rest of it is for.

GLM 5.3 shipped today, and the stack shipped with it: a DFlash 2 drafter, an NVFP4 checkpoint, and a live endpoint on Inco Engine. Together they serve the model at up to 4.4× the throughput of the native FP8 checkpoint with autoregressive decoding, measured at concurrency 1.

If you want your model served like this, fine-tunes included, write to us: contact@inco.ai.

Decode newsletter

Decode with us.

Stay in the loop about research, releases, and ideas about the frontier of efficient inference

One email when we publish something new. Unsubscribe anytime.