Back to News
SourceMarsBit

AMD发布全开源MoE大模型Instella-MoE,16B参数规模挑战主流开源模型

火星财经消息,7 月 25 日,AMD 宣布推出全开源混合专家模型(MoE)Instella-MoE,该模型拥有 160 亿总参数,每个 Token 激活 28 亿参数,号称在同规模开源语言模型中具备领先性能。 AMD 表示,Instella-MoE 完全基于自家 AMD Instinct MI300X 和 MI325X GPU,以及 ROCm 软件栈从零训练完成,并采用 Gated Multi-head Latent Attention(Gated MLA)和 FarSkip-Collective 等架构创新,以提升训练和推理效率。 性能测试显示,Instella-MoE-16B-A3B 基础模型平均得分达到 76.7 分,在全开源模型中排名领先,超过 SmolLM3-3B、OLMo-3-7B 等模型,并在仅激活 28 亿参数的情况下,与更大规模模型竞争。 此外,该模型支持 64K Token 长上下文处理,并完成预训练、中期训练、长上下文扩展、监督微调(SFT)、直接偏好优化(DPO)以及强化学习(RL)等完整训练流程。 AMD 此次同步公开 Instella-MoE 全部模型权重、训练配置、数据配比、中间检查点及推理代码,推动开放 AI 模型研究与复现。 AMD 表示,Instella-MoE 展示了基于 AMD 硬件和开放软件生态进行大规模 MoE 模型训练的能力,未来将继续推进更大规模、更强推理能力和更高效率的开源语言模型研发。
Disclaimer: The views above are the author's only and do not represent 711BTC. Nothing here constitutes investment advice.

Related

06-27 11:08Important

Opinion: The widespread adoption of inexpensive open-source models has led to a decline in the spot rental price of H100 GPUs and a drop in the token price index.

According to Mars Finance, on June 27, research firm Silicon Data released data indicating that the widespread adoption of inexpensive open-source models has led to a decline in the spot rental price of H100 GPUs and a drop in the token price index. However, the firm believes that this does not represent a weakening of overall AI demand. It points out that the large-scale shift of users to inexpensive open-source models has actually increased overall computing consumption, maintaining the demand for high-end computing resources.

07-06 23:27

The ENS governance crisis has escalated, and the community plans to propose delegating 5 million ENS tokens to reform the allocation of voting rights.

According to Foresight News , AvsA, a core contributor to the ENS DAO, recently released a draft proposal on the forum, suggesting reforming the DAO governance structure by delegating (rather than distributing) 5 million ENS tokens. According to the draft, the 5 million ENS tokens will be transferred from the DAO treasury to a multi-delegation contract, with 1 million tokens allocated to each of five stakeholder groups (users, integrators, developers, traditional domain name system participants, and the governance community). A maximum of 10 candidates from each group will be selected based on criteria. The delegated parties will not receive any financial rights in the tokens, only voting rights. If a party does not participate in voting for more than 6 months, their voting rights will be redistributed. The proposal is based on the fact that ENS governance is in crisis, with a single proxy holding enough voting power to pass any proposal independently, and total delegated votes continuing to decline, resulting in significantly low participation in recent proposal voting. This proposal is currently in the feedback collection phase and has not yet entered the formal voting process.

07-06 14:41Important

More severe than the dot-com bubble: Token consumption plummeted by 20%, and the gap between AI investment and sales growth reached 46%.

According to Beating's monitoring, the Silicon Data LLM Token consumption index, which tracks users' actual computing power expenditure, has fallen nearly 20% from its May high. This sudden halt in high growth sends a crucial warning to investors: large model vendors may be losing pricing power with cost-sensitive clients, and it has also raised doubts about the ultimate return on investment for the hundreds of billions of dollars in AI capital expenditure. The divide between bulls and bears has intensified. Bears point out that Allianz Research data shows the growth gap between AI investment and sales has reached 46%, exceeding the 32% imbalance seen during the 2001 telecom bubble burst. Bulls counter that while the average token price has plummeted by 90% since 2023, total expenditure has still nearly doubled, meaning the index decline is merely a structural digestion after price cuts stimulated consumption, and the long-term return on investment in the inference phase is far more optimistic than in the training phase. Increased policy regulation is translating into hidden compliance costs for enterprise users. Washington has imposed stronger policy scrutiny on the distribution and cross-border access of cutting-edge models (such as the release review of OpenAI and geopolitical export controls on Anthropic models). Coupled with the EU's Artificial Intelligence Act's stringent compliance requirements for top-tier models, this has placed a heavy policy burden on leading platforms. To mitigate geopolitical and compliance risks, corporate CFOs have a more rational reason to proactively shift their workloads towards lightweight models that are less subject to regulatory constraints. Subtle changes are also emerging in the hardware chip sector. Although orders for top-tier GPUs and high-bandwidth memory (HBM) are booked until 2026, with substantial supply-demand easing not expected until 2028, the market's main procurement focus has shifted from training chips to inference optimization hardware, and the winners are being reshuffled.

07-03 18:02

JPMorgan Chase: Large-scale model usage and GPU leasing prices rise simultaneously, continuing to support AI infrastructure demand.

According to Mars Finance, JPMorgan Chase's latest "Data Center Watch" report shows that in June, large model usage, API spending, and non-cloud vendor GPU leasing prices all strengthened, indicating that demand for AI infrastructure is still expanding. Although model token prices continued to decline year-on-year, the increase in usage has significantly offset the impact of price reductions, and the overall unit economics of model providers are showing an improving trend. (Cailian Press)

06-26 16:33Important

US AI chip stocks fell across the board in pre-market trading, with AMD dropping nearly 3%.

According to data from Odaily Odaily, AI chip stocks in the US stock market fell collectively in pre-market trading. Intel, Lattice Semiconductor, and AMD fell nearly 3%, while NXP, Qualcomm, Broadcom, Nvidia, TSMC, and Tesla fell more than 1%. Microsoft rose 1%, Amazon rose 0.5%, and IBM rose 0.4%. MSX is a leading RWA trading platform that has listed hundreds of RWA tokens, including popular US stock and ETF tokens such as NVDA, GOOGL, MSFT, AMZN, META, TSM, and AMD.

08-14 19:23

Shinhan Asset Management partners with Plume on tokenized fund pilot

Shinhan Asset Management and Plume will test a tokenized fund using a Korean won-denominated ultra-short-term bond fund as its underlying asset.