Independent Researcher Speech · Image · Video · LLM multimodal generative models Research→Infrastructure→Built AI algo serving 250M+ users 公众号: Generative AI
Singapore
Joined November 2016
- Tweets2.7K
- Following1.5K
- Followers4K
- Likes1.1K
Pinned Tweet
Qwen3.8-27B below 3 bpw: we’ve released two GGUF versions at 2.96 / 2.48 bpw, weighing in at just 9.71 / 8.20 GiB, both with MTP included. Both achieved 100% on AIME 2026 in our evaluation.
We’ve also slimmed down the vision encoder and resolved its GGUF quantization compatibility issues—no modifications to llama.cpp required. The Q4_K version is just 314.46 MiB, 64.59% smaller than BF16, while retaining 0.99537 cosine similarity to the original HF FP32 vision outputs.
Quant 现在也是热闹起来了
好了不藏着了：我也构建一个量化方法比你俩都好 LOL
We’re releasing new DeepSeek-V4.1-Flash-1017 GGUFs, ranking highest across benchmarks.
Dynamic V3.1 leads across nearly all benchmarks: GPQA Diamond, AIME25, LiveCodeBench v6, MMLU-Pro, Aider & more.
Run on 64GB+ RAM.
GGUF: huggingface.co/unsloth/DeepS…
Blog: unsloth.ai/docs/models/deeps…
Pika H3 Flex 基于 MiniMax H3 的后训练与推理优化：把音视频生成从50步压到6步，补画质、偏好对齐和编辑能力。
1. 先提质量，再蒸馏。 用自建美学数据微调 H3；普通 DMD 训练不稳定，改用 PDMD，得到6步音视频联合生成模型。
2. 蒸馏后再对齐。 分别训练 ReFL、DPO 两个 LoRA，改善清晰度、美感和指令遵循，合并后不增加生成步数。
3. 优化推理系统。 缓存 AdaLN 投影，结合融合算子、FlashAttention-4、块稀疏注意力和通信压缩。在 8张B300、5秒768p带音频视频上，将固定6步模型的推理耗时从3.409秒降到2.129秒。
4. 把后期处理接进流程。 原生生成480p/768p，再用单步超分模型输出最高4K；五条音轨则由混合音频后分离得到。（这里卖自家 Pika API）
评测显示画面和音频质量改善，但物理、多场景叙事与复杂参考控制仍有短板。
github.com/Pika-Labs/Pika-H3…
Feiteng retweeted
☎️ Calling all students:
You can now access hundreds of textbooks directly in Notebook, for no charge. From publishers like S Chand Publishing, OpenStax, and Open Textbook Library— browse, select and upload them as sources to ace all of your upcoming exams (no more excuses, ok?)
可以在 EdgeSpeak 里免费使用 YuE2 3B 生成歌曲，质量比肩 Suno v4.5，无需购买。
模型也开源出来了，可以独立部署。
输入歌词，用文字描述曲风、人声和配器，也可以自己写 ABC 乐谱控制旋律与和弦。
这次我优化了采样方法和 GGUF 量化：
· 修改官方 midpoint diffusion 采样策略为 AB3：14 NF vs midpoint 64 NF，生成音频更清晰
· 量化对照 BF16：潜变量余弦相似度 0.99968，Top-1 一致率 95.0%
· YuE2 BF16 7.26 GB → Q5_K 2.57 GB，体积减少 64.6%
· VAE FP32 265.5 MB → Q6_K 59.7 MB，体积减少 77.5%
EdgeSpeak：edgespeak.com
GGUF：huggingface.co/LattifAI/YuE2…
#EdgeSpeak
只有我一个人 领不到吗？？
chatgpt.com/settings/billing 里面没看到领取呀
是 Google Play 付款的问题吗
在 DGX Spark 大幅度涨价前三天想着入手一台玩玩，但想了想家里还有三张 NV 卡吃灰呢就算了吧没时间玩
当时买来当理财也很好呀🤢
苹果 512GB 内存的 Mac Studio 开卖了。M5 Ultra 顶配加 512GB，起步 $22,299，光内存这一项就比 96GB 的基础款贵 $15,500。
同样凑 512GB，买两台 256GB 的 M5 Ultra 是 $21,598，还便宜 $701。
我两台 DGX Spark 加起来 256GB，当初 $4,499 一台买的，一共 $8,998。现在 Micro Center 最低 $6,949 一台，两台就是 $13,898。
按每 GB 内存的价格算：
1. 512GB Mac Studio 约 $43.6
2. 两台 256GB Mac Studio 约 $42.2
3. 两台 DGX Spark 按现价约 $54.3
4. 两台 DGX Spark 按我买的价约 $35.1
带宽差得更多。M5 Ultra 是 1.2TB/s，一台 Spark 是 273GB/s，单路生成 Mac 会快不少。
按现价算，Spark 每 GB 已经比 Mac 还贵了，我这两台算是赶在涨价前买到的。
在我的 DeepSeek-V4.1-Flash-1017 快速测试里：@UnslothAI 的所有 Metrics 仍然优于 ISTA GSQ-RCO
严谨的角度来说，还应该增加 benchmark 的结果 并关注推理 TOKEN 的智能密度
Tibo Day 25: Pro 100、Pro 200、Pro 500 三档和 Business Premium 的用量 x2
拭目以待：下个新模型价格是否会 x2 🤪🤪
Day 25/
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.
真的可以 再也不用担心 Claude 账号被封 设计系统和资产丢失了
发布 baoyu-design v2，完整的把 Claude Design、Claude Slides 和 Claude Dashboard 移植过来了，和原版体验几乎一致（除了没有评论功能），可以运行在任何 Agent 任何模型，不再依赖于 Claude Artifacts。
完全开源：github.com/jimliu/baoyu-desi…
Feiteng retweeted
llama.cpp
llama.cpp v0.8.0 adds LiquidAI LFM3 support across text, vision, audio and omni models.
The new experimental mtmd streaming audio API is joined by server endpoints /v1/realtime and /tts for realtime audio input and output.
Sparse attention for QSA and DSA now uses a fused masking op.
A new SLRU cache manages MoE experts stored in system RAM with an even split.
DeepSeek V4.1 Flash support also lands, along with the ggml v0.28.0 update.
github.com/ggml-org/llama.cp…
🤖 Made with AI
H-Net：直接从原始字节学习，不依赖预先设计的 tokenizer
受控 scaling law 实验中，效率提高 3–5倍。
更大规模实验中，用约 1/10的数据和计算量，达到一个训练了 11T tokens 的开源模型的水平。
在代码上的 Token 也更合理了，我猜测 Claude 应该也做了类似的 tokenizer 优化
I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data.
Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less?
Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.
NVIDIA Nemotron 3.5 Transcribe，语音 encoder 接 LLM，说话人信息提前注入，时间戳用 CTC 对齐。 MoE 结构也是现在多任务音频理解经典配方，总参数 34B 激活 4.3B。
600M ASR encoder 每 80ms 输出一帧，Sortformer 提供说话人信息。音频按 30 秒编码后拼接，长程依赖交给 LLM。
Congrats to @PiotrZelasko and the team!
NVIDIA’s best speech model to date: Nemotron 3.5 Transcribe! It’s an offline model with:
🤖SpeechLLM backed by Nemotron 3.5 Lightning
🧠Steered, promptable, contextual transcription
🌏40 languages
🗣️Up to 8 speaker transcription
⏰Timestamps for each word and speaker
🔥over 2000 RTFx for NVFP4 + DFlash2 on Blackwell GPUs
🏆#1 open model on Open ASR, tied #1 overall on AA-WER, #1 open on David AI multispeaker ASR
🔓Weights, code, recipes, agentic skills, Granary-v2 - all in the open!
See the clip from Her Girl Friday (1940): captioned entirely with a single forward pass!
整理了 #EmbeddingGemma2 开源资源清单：
• DigUp：用自然语言搜索 Mac 上的文档、图片、音频和视频
• Recall：Linux 本地文件搜索
• 微调 Decision 模型：在文本编码器上加一个小决策头，输出选择、布尔判断和评分
• llama.cpp、MLX、LiteRT、Transformers.js：不同设备上的部署路线
欢迎补充 github.com/lattifai/awesome-…
#EmbeddingGemma2 论文发布了：
EmbeddingGemma 2: Unified and Efficient Multimodal Representations
28 页，介绍文本、图片、音频、视频的统一表示与分阶段训练，也讨论按需加载编码器、视觉 token 预算、向量截短和量化等端侧设计。
论文：arxiv.org/abs/2610.37364
资源清单：github.com/lattifai/awesome-…
