Particle 发布 particle-1:面向代理编程、宣称约少用 45% token 的编码模型
Suhail 新公司 Particle 发布 particle-1:基于 DeepSeek V4.1 Flash 后训练,强调代理编程场景下的 token 效率,并已上架 OpenRouter 与 Vercel AI Gateway。
编码代理账单里,输出 token 往往比“模型聪不聪明”更先砸到成本墙上;尤其在多步调试与反复读仓时,省 token 近似于直接省钱,也减轻网关排队压力。
据 @AGTPinsights 对发布信息的整理(2026-10-28),Suhail 的新公司 Particle 推出首个模型 particle-1:在 DeepSeek V4.1 Flash 上后训练,面向调试、规划改动、拆解多步问题等 代理式编程 场景,并宣称在接近基座精度的前提下平均约 少用 45% token。
公开渠道与标价(帖内转述)包括:
- OpenRouter 上架;Particle 称相对 DeepSeek V4.1 Flash,在四项编码基准上输出 token 少约 28%–52%
- Vercel AI Gateway 以
particle/particle-1提供,方便和现有前端/代理流水线对接 - 定价转述为约 $0.25 / 百万 input、$1 / 百万 output(基座非高峰价被对比为 $0.15 / $0.60)
帖内也记录争议点:有观察者指出对照基座可能是较旧的 Flash 版本,同时认可 Particle 披露了基座来源。社区声音(Perplexity、Box、Y Combinator、Vercel 等人士)多落在“把开源权重做得更省 token、更便宜可服务”。
本稿依据二手整理帖与其中引用的官方口径,未独立复现基准数字;选型前仍应以 Particle / OpenRouter / Vercel 页面的模型卡与计价为准,并自行做回归集对比。
Particle launches particle-1 token-efficient coding model
Particle just launched particle-1, a coding model that uses about 45% fewer tokens. Here's what you need to know.
Suhail announced his new company Particle today. particle-1 is its first step in a series of models focused on developers. The focus is token efficiency, and Particle says it keeps parity in performance while using ~45% fewer tokens.
particle-1 is post-trained on DeepSeek V4.1 Flash. It's built for agentic coding work like debugging, planning code changes, and breaking down multi-step problems. Particle says it finishes agent tasks sooner and at lower cost.
It's live now through two partners. OpenRouter lists it, and says it uses 28 to 52% fewer output tokens than DeepSeek V4.1 Flash across four coding benchmarks, per Particle. Vercel added it to AI Gateway as particle/particle-1. Suhail also said the launch includes Particle's own inference service.
PerSplit of reactions: Perplexity's Aravind Srinivas said Particle optimizes open-weight models to be more token-efficient and cheaper to serve. Aaron Levie said agents will consume 1,000X more tokens than today. Garry Tan said Particle will be on the frontier for the smartest model at the lowest dollar per unit of intelligence. Vercel's Guillermo Rauch said Particle's mission is to make intelligence as accessible and abundant as electricity.
One caveat: Teortaxes noted the base is DeepSeek V4.1 Flash 1017, released only 11 days ago. He says Particle seems to compare against the older 0910 version, and credited Suhail for disclosing the base.
Key numbers:
- ~45% fewer tokens on average, accuracy close to the base model
- 28 to 52% fewer output tokens across four coding benchmarks
- $0.25 per million input tokens
- $1 per million output tokens
- Base model: DeepSeek V4.1 Flash, at $0.15 input and $0.60 output off-peak
The Particle-1 launch is a bet that using fewer tokens matters as much as generating them faster, at a time of compute shortages.