Fastest Open-Source LLM APIs We Streamed (2026)

The fastest open-weight LLM API we streamed was DeepSeek V4-Flash on its native api.deepseek.com endpoint: a 0.71 second median time to first token and 95.8 decode tokens per second, measured across three runs on 2026-07-26.

"Fastest open-source LLM" is an ambiguous question, so this page pins it down. It ranks only the open-weight subset (models that publish their trained parameters), and it ranks hosted API speed, not the speed you would get self-hosting on your own GPUs. DeepSeek, Qwen, GLM, Kimi, and MiniMax all ship open weights; GPT-4o is closed and appears here only as a labeled reference row.

An open-weight model is a class of LLM that publishes its trained parameters under a license permitting third parties to download and self-host the model. That is a licensing fact, not a speed fact. The endpoint latency below is what a vendor's own inference stack delivered on 2026-07-26, and it tells you nothing about how fast the same weights would run on hardware you rent yourself.

fastest open source llm (verified 2026-07)

We streamed one fixed prompt ("List 10 numbered one-sentence facts about the HTTP protocol") to each endpoint three times over Server-Sent Events, temperature 0, and report medians. DeepSeek V4-Flash was measured both natively on api.deepseek.com and via OpenRouter routing; every other model was OpenRouter-routed, disclosed, and flagged needs_native_reverify. This is a single-prompt, single-region snapshot, not a load test.

Among the open-weight endpoints, DeepSeek V4-Flash native won both metrics that matter for interactive use: fastest first token at 0.71s and fastest decode at 95.8 tokens per second. GPT-4o's 174 tokens per second is the highest throughput in the whole set, but it is closed, so no open-weight model was fastest on raw decode here.

First-hand evidence: what we measured

We streamed each of the seven endpoints three times on 2026-07-26 and recorded medians directly from the SSE stream. DeepSeek V4-Flash native returned its first token in 0.71s and finished the full 10-item list in a 4.09s median. The same DeepSeek family routed through OpenRouter took 2.54s to first token and 9.68s total, roughly halving decode throughput to 39.5 tokens per second in our runs. Treat that native-versus-routed gap as a routing-overhead signal, not a perfect A/B, because the routed alias may map to a different DeepSeek build or a different physical host.

Kimi K2 posted the fastest usable first token among the routed Chinese open-weight endpoints at a 2.27s median (MiniMax-M2.5's raw 1.50s first token was hidden reasoning, unreadable until 10.68s), ahead of the routed DeepSeek path and well ahead of Qwen. Its decode was modest at 39.4 tokens per second, so it starts quickly but streams the body at a measured pace once it begins.

Qwen3-235B-A22B was the slowest to first token in the entire set at a 3.16s median, even though it is often the cheapest option in our separate task-cost benchmarks. Cheap and fast are different axes, and this run is the plain evidence: the lowest sticker price did not buy the lowest latency on a single-region streaming call.

Reasoning models pay a first-token tax

Two of the open-weight endpoints behaved as reasoning models on this prompt. GLM-5 streamed thinking tokens at a 2.50s TTFT but did not emit a token a user would actually read until a 13.89s median, after roughly 807 reasoning tokens. MiniMax-M2.5 reached first content at 10.68s after about 449 reasoning tokens, despite a quick 1.50s first-token time.

This is a fit finding, not a quality knock. Reasoning overhead can be worth the wait on hard multi-step problems, and it is pure latency tax on a trivial listing task like ours. In an earlier 2026-07-10 non-streaming task benchmark with a shorter output cap, MiniMax-M2.5 posted a 3.64-second median on short tasks, so reasoning-trace depth clearly varies with the prompt and the configuration you send.

Licensing: the open-weight part of "open-source"

The five Chinese families here are open-weight, and their licenses differ enough to matter for anyone planning to self-host. According to DeepSeek's model card on Hugging Face, the weights and repository are released under the MIT License. Qwen sits under a different permissive license: according to Qwen3's model card on Hugging Face, the weights ship under Apache 2.0.

Kimi K2 is published by Moonshot AI under a Modified MIT license, and MiniMax publishes its M-series weights under a Modified MIT license as well. According to MiniMax's platform documentation, the hosted API is the vendor's managed path, which is the surface we timed rather than the downloadable weights. The GLM open-weight family is distributed under the MIT license on its model cards, so all five families cluster around the permissive MIT and Apache families rather than restrictive research-only terms.

Two entity facts anchor the pricing side of the same models. DeepSeek V4-Flash is priced at $0.14 per 1M input tokens on a cache miss, and it is priced at $0.28 per 1M output tokens. According to DeepSeek's API pricing docs, deepseek-chat maps to the non-thinking mode of V4-Flash at exactly those two rates; we last confirmed the billed rate first-hand in our 2026-07-10 runs.

Speed is not quality. This cluster measured latency only, and in our separate 2026-07-10 task benchmark the Chinese models we tested completed short tasks at correctness parity, so the numbers above should drive latency budgeting, not model selection on accuracy grounds.

FAQ

What is the fastest open-source LLM API? In our 2026-07-26 streaming test, DeepSeek V4-Flash on its native api.deepseek.com endpoint was fastest, at a 0.71s median first token and 95.8 decode tokens per second. Every other figure was OpenRouter-routed and needs native re-verification.

Is this the speed I would get self-hosting the open weights? No. We measured hosted vendor APIs, not self-hosted inference. Self-hosted speed depends entirely on your GPUs, batching, quantization, and serving stack, so treat these numbers as a hosted-endpoint snapshot rather than a ceiling for your own deployment.

Which open-weight model started fastest through a router? Kimi K2 posted the fastest usable routed first token among the Chinese open-weight endpoints at a 2.27s median; MiniMax-M2.5's raw 1.50s first token was hidden reasoning. Qwen3-235B-A22B was the slowest to first token at 3.16s, even though it is frequently the cheapest option in cost-focused benchmarks.

Why did GLM-5 and MiniMax-M2.5 look slow to first content? Both behaved as reasoning models on this prompt, streaming thinking tokens before user-visible content. GLM-5 reached first content at a 13.89s median and MiniMax-M2.5 at 10.68s. That is a latency tax on simple generation and can be worth it on harder problems.

Does a lower price mean lower latency? Not here. Qwen3-235B-A22B is often the cheapest per token in our cost benchmarks yet had the slowest first token in this streaming run. Price and latency are independent axes, so budget for each separately.

For the full seven-endpoint leaderboard and the native-versus-routed methodology, see the hub at fastest Chinese LLM API. For how open-weight availability differs across Chinese and Western vendors, see Chinese vs Western open weights, and for token economics see DeepSeek API pricing.

Methodology: SSE streaming, one fixed prompt, temperature 0, max_tokens 4000, three runs per endpoint, medians reported, captured 2026-07-26 from a single network region. DeepSeek native measured on api.deepseek.com; all other endpoints OpenRouter-routed and flagged for native re-verification.

Author: Kevin Fan, Customer Success Manager. Independent editorial analysis for China LLM Directory; no vendor sponsored this benchmark.