Use cases › Speed

Answer before the user notices: low latency and realtime voice

Voice agents, autocomplete and live UI, where the first token has to arrive in a few hundred milliseconds.

20 chat models with a measured first token at these constraints. Data observed 2026-10-25 06:54 UTC.

Showing models for: first token within 300 ms.

What it looks like in practice

DeepSeek-V4.1-Flash sends its first token in 112 ms; the eighth fastest chat model takes 229 ms.

Fastest live endpoint per model, time to first token. Teal rows are this page's picks.

Source: OpenRouter endpoint stats, one snapshot from one client location, observed 2026-10-25

Start with these three

One pick per constraint group, each the leader on its group's rule: the cheapest endpoint that is fast enough, the fastest small-active MoE, the fastest dense model of 32B or fewer. The numbers and dates are the claim; the order is not a verdict.

Cheapest at speed

Qwen3-4B-Thinking-2507

Qwen

Cheapest input price among endpoints with a first token within 300 ms and at least 80 tokens per second: $0.010 per 1M (Nscale, 2026-10-25).

270 msfirst token

161 tok/soutput

$0.010per 1M input

Sources: huggingface.co 2026-10-25

Small active params

Qwen3-30B-A3B

Qwen

Fastest first token among mixture-of-experts chat models with 5B or fewer active parameters: 230 ms (DeepInfra, 2026-10-25).

230 msfirst token

3B of 30.5Bactive

$0.12per 1M input

Sources: huggingface.co 2026-10-25 · huggingface.co 2026-10-25

Dense, mid-size

Qwen3-14B

Qwen

Fastest first token among dense chat models of 32B parameters or fewer: 154 ms (DeepInfra, 2026-10-25).

154 msfirst token

14.8Bparams

$0.070per 1M input

Sources: huggingface.co 2026-10-25 · huggingface.co 2026-10-25

Recipes

How people wire a model into this job. Notes are written by hand from the cited cards and docs; the model links go to the facts.

Prefill-only classification service

Model: Qwen/Qwen3-0.6B
Harness: vLLM with max_tokens=1 and logit readout over option tokens, or an encoder with a classification head
Weights: BF16 (fits edge/single-GPU class)

Score options from logits instead of generating text; no decode phase.
Batch aggressively; prefill is compute-bound and batches well.
Validate on a labeled set; small models drift with prompt wording.

Everything that fits 20

The first ten chat models with a measured first token, in the scoreboard's order for this job. The full list, sortable, lives in the scoreboard with these constraints.

Everything that fits: first 10 of 20
#
1	DeepSeek-V4.1-Flashdeepseek-ai	112 ms	309	$0.020	34	MIT
2	Qwen3-14BQwen	154 ms	104	$0.070	4	Apache 2.0
3	gpt-oss-120bopenai	172 ms	407	$0.030	27	Apache 2.0
4	Qwen3-235B-A22B-Instruct-2507Qwen	200 ms	35	$0.087	7	Apache 2.0
5	Qwen3.8-27BQwen	205 ms	995	$0.020	23	Apache 2.0
6	Inklingthinkingmachines	208 ms	142	$0.95	4	Apache 2.0
7	GLM-5.2zai-org	224 ms	201	$0.034	28	MIT
8	c4ai-command-r7b-12-2024CohereLabs	229 ms	78	—	1	CC BY-NC 4.0
9	Qwen3-30B-A3BQwen	230 ms	134	$0.12	1	Apache 2.0
10	Qwen3.5-122B-A10BQwen	232 ms	112	$0.26	5	Apache 2.0

Show all 20 in the scoreboard →

Most-shared on this job

Links about these models ranked by how many newsletters and link sources shared them in the last 28 days; each source counts once.All links

Shared by 5
DeepSeek-V4.1-Flash-1017 (Hugging Face Repo)Hacker News, The Neuron, TLDR AI, AI Breakfast and 1 moreAbout DeepSeek-V4.1-Flash
huggingface.co · first seen 2026-10-17
Shared by 2
DeepSeek V4.1 Flash: Large-Scale Inference DeploymentLatent Space / AINews, TLDR AIAbout DeepSeek-V4.1-Flash
zhuanlan.zhihu.com · first seen 2026-10-21
Shared by 2
DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made RealThe Neuron, Turing PostAbout DeepSeek-V4.1-Flash
api-docs.deepseek.com · first seen 2026-10-19
Shared by 2
Chinese AI lab DeepSeek updated its V4.1-Flash model, which it says now finishes tasks faster with shorter answers, and…The Rundown AI, Don't Worry About the VaseAbout DeepSeek-V4.1-Flash
x.com · first seen 2026-10-19
Shared by 2
DeepSeek V4.1 Flash 1017Hacker News, AlphaSignalAbout DeepSeek-V4.1-Flash
openrouter.ai · first seen 2026-10-17

How people judge this job

OpenRouter endpoint stats (first token, tokens per second)

None of these measure time to first audio for a voice agent. The first-token times here are text endpoints from OpenRouter's endpoint stats, one snapshot from one client location; a voice loop adds speech recognition and synthesis on top.

Nearest scoreboard view: Low latency.