以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

Zyphra

37 atoms · 跨 11 天 · 首见 2026-04-30 · 最近 2026-06-25

三色: 🟦 fact 24 · 🟥 take 13 · stance ▲13/▼2/◆8

叙事状态: 🍂 fading (退潮) · 💤 沉寂 · 策展近14d 0 atom

状态只算低频策展源 (群/卖方); X firehose 仅作背景音量

来源: X 37

时态: fresh:25 · aging:11 · stale:1

标签: 好信源:19 · 好数字:9 · 好观点:8 · 好思考:6 · 好问题:1

🟨 AI 综合 · junior analyst 概览

展开 AI 综合 (灰色 · 非市场结论 · 点击数字溯源到原 atom)

model: deepseek-chat · 2026-07-12 · 默认折叠

Zyphra 刚以至少 $5B 估值完成 $500M 的 Series B 融资 ¹ ¹. 其最新 ZAYA1-8B 模型以 <1B 激活参数在数学和推理上超越多数开源模型 ¹ ¹. 推理优化性能显著超过 AMD baseline, 缩小了 MI355X 与 B200 的差距 ¹. 模型设计工作和 MoE 路由改进受到认可 ¹ ¹. 但另一方面, 其 MoE 专家规模并未超越 DeepSeek 的已有做法 ¹. 大规模能力尚未得到验证, 而 DeepSeek 已证明自己能构建大规模有效系统 ¹.

🧭 拥挤度 (一人一票): ▲ 4 位作者 (KOL4) vs ▼ 2 位作者 (KOL2)

📄 广泛报道的事实 · 2 个

展开 (多人转述同一事实/数字 — 确认度高, **非独立观点**, 不标多空)

first work in continual learning: Can LLMs learn forever from new data? — 2 源报道 · X
continual learning path to AGI through recursive self-improvement (RSI) — 2 源报道 · X

🟦 客观事实 (facts) (24)

<a id="atom-40962ab91f78c55a"></a>🟦 2026-05-19 X·@Techmeme ⭐ · Series B 融资金额 · 融资
Zyphra is raising a $500M Series B
tags: 好数字·好信源 · → daily
value: qty=$500M · date=2026-05-19
<a id="atom-b7e0f9cc86903ff5"></a>🟦 2026-05-19 X·@Techmeme ⭐ · Series B 估值 · 融资
at a valuation of at least $5B
tags: 好数字·好信源 · → daily
value: qty=$5B · direction=at least · date=2026-05-19
<a id="atom-d54736ff65440e3c"></a>🟦 2026-05-06 X·@teortaxesTex · ZAYA1-8B 发布 · 模型发布
Today we're releasing ZAYA1-8B, a reasoning MoE trained on @AMD and optimized for intelligence density. With <1B active params, it outperforms open-weight models many times its size on math and re
tags: 好信源·好数字 · → daily
value: qty=8B · date=
📷 原图
<a id="atom-a570fe8125bd36f4"></a>🟦 2026-05-20 X·@FutureJurvetson · ZAYA1-8B 训练预算 · 融资情况
ZAYA1-8B is the company's first frontier reasoning model and was trained while operating from a single 100 million dollar Series A round.
tags: 好数字·好信源 · → daily
value: qty=$100M · date=Series A round
📷 原图
<a id="atom-b8d67e6b2c22feab"></a>🟦 2026-04-30 X·@Underfox3 · TSP在测试序列长度上实现最低峰值内存 · 内存效率/训练性能
The empirical evaluation on up to 1024 MI300X GPUs show that TSP achieves the lowest peak memory compared to TP, SP, and TP+SP meshes across all tested sequence lengths while maintaining competitive throughput.
tags: 好数字·好信源 · → daily
📷 原图
<a id="atom-b2fe69f23160fd51"></a>🟦 2026-06-12 X·@ZyphraAI · released ZONOS2 model · 产品发布
Today we're releasing ZONOS2, our next-generation real-time TTS model with high-fidelity voice cloning.
tags: 好信源 · → daily
value: qty= · date=2026-06-12
<a id="atom-1f7fd8c09a8a04e2"></a>🟦 2026-05-06 X·@teortaxesTex · ZAYA1-8B 性能对比 · 模型发布
With <1B active params, it outperforms open-weight models many times its size on math and re
tags: 好数字 · → daily
value: qty=<1B active params · date=
📷 原图
<a id="atom-b466ffaff8218858"></a>🟦 2026-05-07 X·@teortaxesTex · model size: Zyphra model has 74B parameters · 模型参数/架构验证
74B is entirely sufficient to validate their architecture and methodology.
tags: 好数字 · → daily
value: qty=74B
📷 原图
<a id="atom-2282670973889798"></a>🟦 2026-05-06 X·@teortaxesTex · 80B 模型即将发布 · 模型发布
80B upcoming.
tags: 好数字 · → daily
value: qty=80B · date=upcoming
📷 原图
<a id="atom-7e473e7e4008ac56"></a>🟦 2026-06-25 X·@xlr8harder · first work in continual learning: Can LLMs learn forever from new data? · 技术路线/模型发布
Zyphra is sharing our first work in continual learning where we study: Can LLMs learn forever from new data?
tags: 好信源 · → daily
<a id="atom-8b53fa28874b556f"></a>🟦 2026-06-24 X·@teortaxesTex · scaling law for plasticity loss in continuously trained LLMs · 技术路线
Zyphra fits a scaling law for plasticity loss in continuously trained LLMs.
tags: 好信源 · → daily
📷 原图
<a id="atom-fa8e458d9eff66fd"></a>🟦 2026-05-18 X·@ZyphraAI · 发布 MI355X 推理端到端基准测试 · 产品发布/技术路线
We're publishing our first end-to-end benchmarks for Zyphra Inference on @AMD Instinct MI355X.
tags: 好信源 · → daily
📷 原图
<a id="atom-da0650f9e03bc7c6"></a>🟦 2026-05-07 X·@ZyphraAI · released ZAYA1-74B-Preview model · 模型发布/产品发布
Today we're releasing ZAYA1-74B-Preview, a major milestone in scaling pretraining on @AMD.
tags: 好信源 · → daily
📷 原图
<a id="atom-dd05088931d2e86f"></a>🟦 2026-05-06 X·@1a3orn · release ZAYA1-8B model weights · 模型发布
releases Apache weights for a 8b MoE
tags: 好信源 · → daily
value: qty=Apache weights for 8b MoE · date=2026-05-06
<a id="atom-4dea6f12ee4b54d5"></a>🟦 2026-05-04 X·@ZyphraAI · Cloud 发布 · 产品发布
Introducing Zyphra Cloud: A full stack AI platform on AMD.
tags: 好信源 · → daily
📷 原图
<a id="atom-4e32df267b0eef95"></a>🟦 2026-05-04 X·@ZyphraAI · Inference 发布 · 产品发布
Launching today with Zyphra Inference: serverless inference for frontier open-weight models focused on long horizon agentic workloads.
tags: 好信源 · → daily
📷 原图
<a id="atom-3364fd1dab9a0e41"></a>🟦 2026-05-04 X·@ZyphraAI · GPU 供应商 · 合作客户
Powered by @AMD MI355X GPUs on @TensorWave.
tags: 好信源 · → daily
value: qty=AMD MI355X
📷 原图
<a id="atom-80dc0d18fa564cfe"></a>🟦 2026-05-04 X·@AIStockSavvy · launched AI Cloud Platform · 产品发布
Zyphra launches AI Cloud Platform Powered by $AMD AMD Instinct GPUs
tags: 好信源 · → daily
value: qty= · date=2026-05-04
<a id="atom-b808fe5cf6a8d8f6"></a>🟦 2026-05-04 X·@AIStockSavvy · launched serverless inference service · 产品发布
Launch includes 𝐬𝐞𝐫𝐯𝐞𝐫𝐥𝐞𝐬𝐬 𝐢𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 with open-weight models
tags: 好信源 · → daily
value: qty= · date=2026-05-04
<a id="atom-36f90a8de20d7fde"></a>🟦 2026-05-04 X·@AIStockSavvy · targets workflow automation · 技术路线
Targets agentic coding, deep research, and workflow automation
tags: 好信源 · → daily
value: qty= · date=2026-05-04
<a id="atom-3f3041a0bb75ee17"></a>🟦 2026-04-30 X·@Underfox3 [老化中] · 提出并行执行策略 · 技术路线/训练效率
Zyphra researchers have proposed a parallel execution strategy that folds tensor parallelism and sequence parallelism onto a single device axis, improving long-context and memory-constrained model training.
tags: 好信源 · → daily
📷 原图
<a id="atom-bd8d637d684209d9"></a>🟦 2026-04-30 X·@Underfox3 [老化中] · 提出的方法降低内存 · 内存效率/技术路线
By assigning each rank both a weight shard and a sequence shard, the proposed approach reduces both per-device parameter-proportional (parameters, optimizer states, gradients) memory and activation memory.
tags: 好信源 · → daily
📷 原图
<a id="atom-e000284dc9b03482"></a>🟦 2026-06-25 X·@scaling01 · 研究目标 · 技术路线
Zyphra is sharing our first work in continual learning where we study: Can LLMs learn forever from new data?
tags: 好问题 · → daily
📷 原图
<a id="atom-b55185933daed11e"></a>🟦 2026-05-04 X·@teortaxesTex · 新并行策略发布 · 技术路线
Pretty massive contribution from Zyphra – a whole new parallelism strategy. «Tensor and sequence parallelism assigns each rank both a weight shard and a sequence shard, reducing both parameter and activation memory along the same device axis».
tags: 好思考 · → daily
📷 原图

🟥 多头 takes (bullish) (9)

<a id="atom-c0c6e9b025516251"></a>🟥 2026-05-18 X·@ZyphraAI · 推理优化性能超过 AMD baseline · 技术路线/竞争格局
Our inference optimizations strongly outperform the AMD baseline and narrows the gap between MI355X and B200 for serving Kimi K2.6, GLM 5.1, and DeepSeek V3.2
tags: 好数字·好观点 · → daily
📷 原图
<a id="atom-7123aac873b95f75"></a>🟥 2026-06-25 X·@scaling01 · 研究框架 · 技术路线
Many see continual learning as a path to AGI through recursive self-improvement (RSI).
tags: 好思考 · → daily
📷 原图
<a id="atom-1d55f0c22f433074"></a>🟥 2026-05-06 X·@teortaxesTex [老化中] · AMD 合作推测 · 合作客户
if they expand to greater scale, which I think is what their AMD partnership is about, I may uprank them to DeepSeek tier
tags: 好思考 · → daily
📷 原图
<a id="atom-9ea716c0dcf0759e"></a>🟥 2026-05-06 X·@teortaxesTex [老化中] · MoE 路由改进认可 · 技术路线
you know what, you're right. in terms of expert size and count it's a step back but they reshuffle complexity/granularity into a more expressive router. It's a development contrary to the entire post-DSMoE tree so actually I wasn't giving them enough credit
tags: 好思考 · → daily
<a id="atom-8aef8aac48464978"></a>🟥 2026-05-06 X·@teortaxesTex [老化中] · 小模型路由优于 DeepSeek 推测 · 技术路线
I suspect that Zyphra's approach is better at least in small models. Normal routers are too dumb, and micro-DSMoE experts too smol
tags: 好思考 · → daily
<a id="atom-db0d325854765e83"></a>🟥 2026-06-25 X·@xlr8harder · continual learning path to AGI through recursive self-improvement (RSI) · 技术路线
Many see continual learning as a path to AGI through recursive self-improvement (RSI).
tags: 好观点 · → daily
<a id="atom-6ae03fc772205303"></a>🟥 2026-05-11 X·@teortaxesTex [老化中] · made AMD relevant again · 合作客户/技术路线
They have done what Geohot has failed to do: made AMD relevant again.
tags: 好观点 · → daily
<a id="atom-afd505540a580e86"></a>🟥 2026-05-07 X·@xlr8harder [老化中] · 模型设计工作 · 技术路线
Zyphra is doing some extremely competent model design work
tags: 好观点 · → daily
<a id="atom-aacc8c7feca0e478"></a>🟥 2026-05-06 X·@teortaxesTex [老旧] · SoTA 架构评述 · 技术路线
Their SoTA architecture ('DSMoE-MLA++'), high-end RL and test-time scaling. Frontier open lab, talented and brave.
tags: 好观点 · → daily
📷 原图

🟥 空头 takes (bearish) (2)

<a id="atom-63d14b522890893d"></a>🟥 2026-05-06 X·@eliebakouch [老化中] · MoE 专家规模对比 DeepSeek · 技术路线
fair for attention ig, but for MoE this is not far beyond deepseek, this is literally how people did it before
tags: 好观点 · → daily
<a id="atom-3a116dccd67bf024"></a>🟥 2026-05-06 X·@teortaxesTex [老化中] · 无法证明大规模能力 · 供给产能
DeepSeek builds things that work at huge scale Zyphra is yet to demonstrate that they can do that
tags: 好观点 · → daily

🟥 中性 takes (neutral) (1)

<a id="atom-edd2e80a1bceec0f"></a>🟥 2026-05-06 X·@Dorialexander [老化中] · release · 模型发布/技术路线
Strong new release from Zyphra relying on hybrid synthetic pretraining — standard crawl interspesed with heavy thinking traces throughout training.
tags: 好思考 · → daily
📷 原图

⏱ 时间轴 (近 20)

🟦 2026-06-25 · fact · first work in continual learning: Can LLMs learn forever fro · → daily
🟥 2026-06-25 · narrative · continual learning path to AGI through recursive self-improv · → daily
🟦 2026-06-25 · narrative · 研究目标 · → daily
🟥 2026-06-25 · narrative · 研究框架 · → daily
🟦 2026-06-24 · fact · scaling law for plasticity loss in continuously trained LLMs · → daily
🟦 2026-06-12 · fact · released ZONOS2 model · → daily
🟦 2026-05-20 · fact · ZAYA1-8B 训练预算 · → daily
🟦 2026-05-19 · fact · Series B 融资金额 · → daily
🟦 2026-05-19 · fact · Series B 估值 · → daily
🟦 2026-05-18 · fact · 发布 MI355X 推理端到端基准测试 · → daily
🟥 2026-05-18 · fact · 推理优化性能超过 AMD baseline · → daily
🟥 2026-05-11 · narrative · made AMD relevant again · → daily
🟥 2026-05-07 · narrative · 模型设计工作 · → daily
🟥 2026-05-07 · narrative · 值得关注 · → daily
🟦 2026-05-07 · fact · released ZAYA1-74B-Preview model · → daily
🟦 2026-05-07 · fact · model size: Zyphra model has 74B parameters · → daily
🟦 2026-05-06 · fact · release ZAYA1-8B model weights · → daily
🟦 2026-05-06 · fact · ZAYA1-8B 发布 · → daily
🟦 2026-05-06 · fact · ZAYA1-8B 性能对比 · → daily
🟥 2026-05-06 · position · SoTA 架构评述 · → daily

← 实体目录 · 系统日志