Inco AI

Inco AI is an AI inference company that builds speculative decoding draft models, quantized checkpoints, a hosted inference platform and an open-source local inference engine for Macs. Its public identity is the tagline "inference, reimagined for the agentic era," and its X account gives its location as Palo Alto, California.^[1][2] The company's technology centers on DFlash, a block-diffusion drafting method for speculative decoding that Inco says its team released in January 2026, and on DFlash 2, the successor Inco published on 18 August 2026.^[3][4] In September 2026 it opened a public beta of the Inco platform, which it says was the fastest provider on Artificial Analysis for several open models, and released Splash, an Apple silicon engine that LM Studio integrated at launch.^[5][6][7]

Inco's blog posts and citations are credited to "Inco AI" as a collective author. Zhijian Liu, an assistant professor at UC San Diego, is described as a co-founder of Inco AI in the author biography of a June 2026 NVIDIA Technical Blog post he co-wrote, and his X profile says he is "building something new (@inco_ai)."^[8][21]

People and origins

Zhijian Liu directs Z Lab at UC San Diego, which is part of the university's ML Systems Group and focuses on making AI "smaller, faster, and more efficient."^[9][10] He received his PhD from MIT, advised by Song Han, and has worked as a research scientist at NVIDIA.^[9][8] Z Lab's projects include DFlash (Jian Chen, Yesheng Liang and Zhijian Liu, ICML 2026), ParoQuant (Yesheng Liang, Haisheng Chen, Zihan Zhang, Song Han and Zhijian Liu, ICLR 2026), and several efficient vision-language-action and vision-language model papers.^[10]

The DFlash code repository, z-lab/dflash on GitHub, was created on 4 January 2026 under the MIT license, and Z Lab's news feed lists "DFlash is released!" in January 2026.^[11][9] The paper, "DFlash: Block Diffusion for Flash Speculative Decoding," was posted to arXiv on 5 February 2026.^[12] Inco's August 2026 post says: "Our team released DFlash in January." The same DFlash 2 drafters also appear under Z Lab's Hugging Face organization as z-lab/Muse-Glimmer-30B-DFlash2 and z-lab/Qwen3.8-27B-DFlash2, created on 14 and 15 August 2026, a few days before the incoai repositories.^[13][15]

Inco's GitHub organization was created on 6 May 2026, the same day as its X account.^[2][14] Its Hugging Face organization, incoai, is verified and listed three members in September 2026.^[15]

Technology

DFlash

DFlash replaces the autoregressive draft model in speculative decoding with a lightweight block-diffusion model. The drafter produces a whole block of draft tokens in a single forward pass, conditioned on context features extracted from the target model, and the target then verifies the block in parallel. The paper reports "over 6x lossless acceleration" across a range of models and tasks and up to 2.5 times the speedup of EAGLE-3.^[12] It is covered in more depth in the DFlash article.

Inco's DFlash 2 post makes several adoption claims for the original method: that it runs in SGLang, vLLM, TensorRT-LLM and llama.cpp; that NVIDIA measured up to 15 times throughput with it on Blackwell GPUs and Google reported three times more tokens per second on TPUs; that CoreWeave's Kimi K2.7-Code endpoint runs DFlash by default; that NVIDIA, Red Hat and Modal have published DFlash drafters; and that Meta (Muse Glimmer), Poolside (Laguna), Xiaomi (MiMo-V2.5-Pro) and NVIDIA (Nemotron 3.5 Lightning) ship official DFlash drafters with their models.^[3] It put Hugging Face downloads of DFlash models at more than 3.5 million as of August 2026; the Splash post a month later said "more than 6 million."^[3][6] These are the company's own figures.

DFlash 2

DFlash 2 keeps DFlash's one-pass parallel draft and adds two parts.^[3]

• A path selector. Inco found that the correct token is usually among DFlash's top candidates even when the top pick is wrong: on a five-layer Qwen3-4B drafter on GSM8K, the first position's top pick was right 85.4 percent of the time, but the right token was in the top 16 candidates 99.5 percent of the time. DFlash 2 keeps the top 16 candidates per position and scores every adjacent pair with DFlash's own logit plus a gated low-rank bilinear term over 256-dimensional token embeddings, then walks the best path. On that drafter it raised acceptance length from 4.27 to 4.61 at temperature 0, for 2.0 million added parameters and 0.6 percent added cycle latency.
• A short dynamic convolution. To counter what Inco calls "suffix decay" (accuracy falling toward the end of a draft block), it inserts a two-tap dynamic depthwise convolution before and after each attention and feed-forward sublayer. Inco reports that this added 16.5 million parameters (3 percent) and 0.7 percent latency, and brought a five-layer drafter close to a 15-layer one.

Together the two parts add 1.3 percent to draft-verify cycle latency, according to Inco.^[3] Its reported mean acceptance lengths (tokens per verification step, including the verifier's next token) across GSM8K, MATH-500, HumanEval, MBPP and MT-Bench were:^[3]

Inco translates these into 2.7 to 3.4 times the throughput of autoregressive decoding on Qwen3.8-27B and 3.1 to 4.6 times on Muse Glimmer, with the per-task breakdowns on the model cards.^[3] Inco describes the method as lossless: rejection sampling restores the target model's exact distribution.^[3]

Inco platform and Inco Engine

Inco describes its data-center stack in four layers: fleet-scale serving (orchestration, cache-aware routing and cache management), speculative decoding with DFlash and DFlash 2, model quantization, and "kernel and runtime optimization agents" that adapt low-level execution to each model and hardware target inside what it calls Inco Engine.^[5] The platform post says Inco is "the team behind industry-leading quantization research, including AWQ and ParoQuant." ParoQuant's authors are listed above; AWQ was published in 2023 by Ji Lin, Jiaming Tang and eight co-authors with Song Han as last author, and the post does not name which Inco staff worked on it.^[5][16]

The platform entered public beta on 3 September 2026 with endpoints for Kimi K3, MiniMax M3, GLM-5.3 and GLM-5.3-Flash. Inco's table of Artificial Analysis output speeds, captured on 8 September 2026, was:^[5]

Provider rankings on Artificial Analysis change as providers are added or re-measured. On 23 September 2026 the Artificial Analysis provider pages still listed Inco's endpoint as the fastest for all five models, though several margins had narrowed; Kimi K3, for example, showed 284.7 tokens per second for Inco against 249.9 for Nebius.^[22] The Splash post of 17 September named a fifth model, DeepSeek V4.1-Flash, among those for which Inco said it was the fastest provider.^[6]

GLM-5.3 day-zero launch

Inco was a day-zero serving partner for the open-weight release of Z.ai's GLM-5.3; Inco's post dated 28 August 2026 says the model "arrives today" and links the zai-org/GLM-5.3 checkpoint on Hugging Face. It published a DFlash 2 drafter and an NVFP4 checkpoint and served the model from compute supplied by TokenRouter, on whose Blackwell clusters it trained the drafter and produced the NVFP4 checkpoint.^[17] Inco reported that the two checkpoints plus Inco Engine delivered up to 4.4 times the throughput of the native FP8 checkpoint with autoregressive decoding at concurrency 1. It gave DFlash 2 acceptance lengths of 6.02 on MATH-500, 5.94 on GSM8K and 5.48 on HumanEval, against 5.05, 5.12 and 4.85 for GLM-5.3's native MTP path, and 383.3 output tokens per second on MATH-500 at concurrency 1.^[17] Its NVFP4 accuracy table was within about one point of FP8 on each benchmark it listed (for example GPQA Diamond 91.2 against 91.1, and HLE 35.2 against 35.9).^[17] Preview list pricing was $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens, with a 1M-token context window.^[17]

Splash

Splash, released on 18 September 2026 under Apache-2.0, is a local inference engine for Apple M3 or newer Macs that ships each supported model (Qwen3.8-27B and Qwen3.6-35B-A3B at launch) with its own DFlash 2 draft, generated Metal kernels and a precomputed memory plan.^[6][18] In Inco's measurements on a 48 GB M5 Pro, it decoded Qwen3.8-27B at 74 tokens per second on short prompts (2.0 times the next-fastest engine Inco tested) and reached 170 tokens per second combined across four concurrent requests; Inco's launch post on X cited 144 tokens per second on an M5 Max.^[6][19] LM Studio added Splash as an experimental backend in LM Studio Bionic 1.1.5 and LM Studio 0.4.25 on 19 September 2026.^[7][20]

Hugging Face releases

Source: Hugging Face model API, queried 23 September 2026.^[15] The Qwen and Muse Glimmer drafters carry Apache-2.0, while both GLM drafters are tagged with a non-commercial, no-derivatives Creative Commons license.^[15] Inco's DFlash 2 post gives install commands for SGLang, vLLM, llama.cpp, Ollama and oMLX, several of which build from pull-request branches.^[3]

Timeline

References

1. ^Inco AI. Homepage. inco.ai (accessed September 23, 2026)
2. ^1 ^2 ^3Inco AI (@inco_ai). X profile (joined May 6, 2026; location Palo Alto, CA), via the fxtwitter API. x.com/inco_ai (accessed September 23, 2026)
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Inco AI. "DFlash 2: Keep Drafting Parallel." Inco AI blog, August 18, 2026. inco.ai/...dflash2
4. ^Inco AI. "Decode" blog index. inco.ai/blog (accessed September 23, 2026)
5. ^1 ^2 ^3 ^4 ^5Inco AI. "Inco AI Launches Its Inference Platform, Leading Across Four Open Models on Artificial Analysis." Inco AI blog, September 3, 2026, updated September 8, 2026. inco.ai/...inco-platform-aa
6. ^1 ^2 ^3 ^4 ^5 ^6Inco AI. "Splash: A Local Engine Built Around the Model." Inco AI blog, dated September 17, 2026. inco.ai/...splash
7. ^1 ^2 ^3LM Studio Team. "Splash Engine - the fastest local Qwen3.8 on Apple Silicon." LM Studio Blog, September 18, 2026. lmstudio.ai/...splash-engine
8. ^1 ^2Zhijian Liu (@zhijianliu_). X profile, via the fxtwitter API. x.com/zhijianliu_ (accessed September 23, 2026)
9. ^1 ^2 ^3Zhijian Liu. Personal homepage. zhijianliu.com (accessed September 23, 2026)
10. ^1 ^2Z Lab, UC San Diego. Homepage and projects list. z-lab.ai/projects (accessed September 23, 2026)
11. ^1 ^2z-lab. "dflash" repository metadata (created January 4, 2026; MIT license). GitHub. github.com/...dflash
12. ^1 ^2 ^3Jian Chen, Yesheng Liang, and Zhijian Liu. "DFlash: Block Diffusion for Flash Speculative Decoding." arXiv:2602.06036, February 5, 2026 (v2 May 28, 2026; ICML 2026). arxiv.org/...2602.06036
13. ^Hugging Face. Models by author z-lab (API listing). huggingface.co/z-lab (accessed September 23, 2026)
14. ^1 ^2GitHub. Inco AI organization (created May 6, 2026). github.com/incoai
15. ^1 ^2 ^3 ^4Hugging Face. incoai organization overview and model API listing. huggingface.co/incoai (accessed September 23, 2026)
16. ^Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. "AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration." arXiv:2306.00978, June 1, 2023. arxiv.org/...2306.00978
17. ^1 ^2 ^3 ^4 ^5Inco AI. "Inco AI launches Day-0 support for GLM 5.3." Inco AI blog, August 28, 2026. inco.ai/...glm-5-3
18. ^1 ^2Inco AI. incoai/splash repository and releases. GitHub. github.com/...splash
19. ^Inco AI (@inco_ai). Post on X announcing Splash, September 19, 2026 (00:07 UTC). x.com/...2101100749623341513
20. ^1 ^2LM Studio. "Bionic 1.1.5" release notes, September 19, 2026. lmstudio.ai/...bionic-v1.1.5
21. ^Amr Elmeleegy, Benjamin Chislett, Fernando Xiong, Michael Iovine, Omri Almog, Hao Zhang, and Zhijian Liu. "Boost Inference Performance up to 15x on NVIDIA Blackwell Using DFlash Speculative Decoding." NVIDIA Technical Blog, June 23, 2026. developer.nvidia.com/...flash-speculative-decoding
22. ^Artificial Analysis. Provider pages for Kimi K3, MiniMax M3, GLM-5.3, GLM-5.3-Flash and DeepSeek V4.1 Flash (accessed September 23, 2026). artificialanalysis.ai