OzTianlu (Zixi "Oz" Li)

reacted to their post with 👀 20 days ago

Post

96

zixi-51M: Not a stock voice. A voice. 🐾

NoesisLab/zixi-51m
Everyone is shipping 500B-parameter models. I shipped 51 million. With one voice. On purpose. 🎤

What: open-weights Chinese TTS. VITS2 lineage. 51.43M params. Runs on your laptop CPU — no cloud, no API key, no GPU bill, no meter running.💸

The voice: not a stock voice, not a 5,000-speaker average. One real human's voice — mine. Timbre, rhythm, breath included. The sample on themodel card is one take, default parameters, recorded by absolutely no one. 🎧

What's NOT inside:

• ❌ BERT (zero. the text encoder vibes on phones + tones alone)
• ❌ language routing (Chinese-exclusive: 77 symbols, 6 tones, one job)
• ❌ my training data (the corpus is my own writing and stays mine — open weights, closed corpus, that's the deal 🤝)

The trick: embedding transplant. Phone, tone and speaker rows inherited from the MeloTTS-Chinese base by symbol name — no random init. Themodel woke up already knowing Chinese. 🧠

Built solo, corpus to packaging, by a writer who is also a violinist. Prosody is my day job. 🎻

Use it, fork it, fine-tune your own voice. Just don't put words in a real person's mouth. 🙏

🔗 Weights: NoesisLab/zixi-51m
💻 Code: https://github.com/lizixi-0x2F/zixi-51M

posted an update 20 days ago

Post

96

zixi-51M: Not a stock voice. A voice. 🐾

NoesisLab/zixi-51m
Everyone is shipping 500B-parameter models. I shipped 51 million. With one voice. On purpose. 🎤

What: open-weights Chinese TTS. VITS2 lineage. 51.43M params. Runs on your laptop CPU — no cloud, no API key, no GPU bill, no meter running.💸

The voice: not a stock voice, not a 5,000-speaker average. One real human's voice — mine. Timbre, rhythm, breath included. The sample on themodel card is one take, default parameters, recorded by absolutely no one. 🎧

What's NOT inside:

• ❌ BERT (zero. the text encoder vibes on phones + tones alone)
• ❌ language routing (Chinese-exclusive: 77 symbols, 6 tones, one job)
• ❌ my training data (the corpus is my own writing and stays mine — open weights, closed corpus, that's the deal 🤝)

The trick: embedding transplant. Phone, tone and speaker rows inherited from the MeloTTS-Chinese base by symbol name — no random init. Themodel woke up already knowing Chinese. 🧠

Built solo, corpus to packaging, by a writer who is also a violinist. Prosody is my day job. 🎻

Use it, fork it, fine-tune your own voice. Just don't put words in a real person's mouth. 🙏

🔗 Weights: NoesisLab/zixi-51m
💻 Code: https://github.com/lizixi-0x2F/zixi-51M

reacted to their post with 🔥 5 months ago

Post

6,348

ResNet is Explicit Euler. GPT is Implicit Euler. What Else is Hiding in Plain Sight?

Read online: https://datawhalechina.github.io/learning-terrain/

I wrote an open-source monograph on learning dynamics — The Terrain of Learning. Bilingual (Chinese/English), 4 volumes, 12 chapters, 30+ print-grade figures. Completely free (CC BY-NC-SA 4.0).

The core argument: gradient descent is not optimization. It's terrain motion. The loss function is a landscape. The gradient is the direction of slope. The optimizer is how you choose each step. Once you see it this way, everything clicks:

ResNet = explicit Euler integration on a vector field. The residual branch is the vector field. Each layer takes one Euler step.

GPT autoregression = implicit-state Euler iteration. Stable where explicit Euler explodes. That's why transformers handle long-range dependencies.

DEQ = the Banach fixed-point theorem in production. The forward pass is root-finding. There are no layers to backprop through.

KL divergence = a Bregman divergence on the entropy landscape. Your belief space is curved, not flat.

Chain-of-thought reasoning = hidden states flowing along a reasoning field toward an attractor basin. Correct answers have wide basins. The number of reasoning steps is determined by the terrain, not by the problem.

Diffusion models = systems flowing downhill along a score vector field, from noise to structure, from high energy to low energy.

The book traces one idea across 337 years — from F=ma (Newton, 1687) to H=T+V (Hamilton, 1833) to loss landscape + gradient field (2020s). Hamilton replaced a catalog of forces with one geometric object. This book does the same for deep learning.

GitHub: https://github.com/datawhalechina/learning-terrain
Discussion: https://github.com/datawhalechina/learning-terrain/discussions/2

Convergence is not hope. Convergence is geometry. You see.

1 reply

·

posted an update 5 months ago

Post

6,348

ResNet is Explicit Euler. GPT is Implicit Euler. What Else is Hiding in Plain Sight?

Read online: https://datawhalechina.github.io/learning-terrain/

I wrote an open-source monograph on learning dynamics — The Terrain of Learning. Bilingual (Chinese/English), 4 volumes, 12 chapters, 30+ print-grade figures. Completely free (CC BY-NC-SA 4.0).

The core argument: gradient descent is not optimization. It's terrain motion. The loss function is a landscape. The gradient is the direction of slope. The optimizer is how you choose each step. Once you see it this way, everything clicks:

ResNet = explicit Euler integration on a vector field. The residual branch is the vector field. Each layer takes one Euler step.

GPT autoregression = implicit-state Euler iteration. Stable where explicit Euler explodes. That's why transformers handle long-range dependencies.

DEQ = the Banach fixed-point theorem in production. The forward pass is root-finding. There are no layers to backprop through.

KL divergence = a Bregman divergence on the entropy landscape. Your belief space is curved, not flat.

Chain-of-thought reasoning = hidden states flowing along a reasoning field toward an attractor basin. Correct answers have wide basins. The number of reasoning steps is determined by the terrain, not by the problem.

Diffusion models = systems flowing downhill along a score vector field, from noise to structure, from high energy to low energy.

The book traces one idea across 337 years — from F=ma (Newton, 1687) to H=T+V (Hamilton, 1833) to loss landscape + gradient field (2020s). Hamilton replaced a catalog of forces with one geometric object. This book does the same for deep learning.

GitHub: https://github.com/datawhalechina/learning-terrain
Discussion: https://github.com/datawhalechina/learning-terrain/discussions/2

Convergence is not hope. Convergence is geometry. You see.

1 reply

·

reacted to their post with 🤗 7 months ago

Post

1,432

https://github.com/lizixi-0x2F/March
I just released March, an open-source high-performance KV cache sharing library for LLM inference that uses Trie-based prefix deduplication.
When you run LLM services, you often see thousands of requests sharing the same system prompt and conversation history. But traditional KV cache systems store each sequence separately — duplicating the exact same data over and over again. Pure waste.
March uses a Trie structure to automatically detect and reuse identical token prefixes. Instead of storing [system_prompt + history] 1000 times, it's stored once. Everyone shares it.
- 80-97% memory reduction in prefix-heavy workloads (tested on SmolLM2-135M with 500 multi-turn conversations)
- Zero-copy queries — returns direct pointers into the memory pool, no expensive memcpy on the hot path
- Predictable memory usage — fixed-size page pool with O(L) complexity
- Trade-off: slightly slower than dict O(1) lookup, but the memory savings are worth it in production

1 reply

·

posted an update 7 months ago

Post

1,432

https://github.com/lizixi-0x2F/March
I just released March, an open-source high-performance KV cache sharing library for LLM inference that uses Trie-based prefix deduplication.
When you run LLM services, you often see thousands of requests sharing the same system prompt and conversation history. But traditional KV cache systems store each sequence separately — duplicating the exact same data over and over again. Pure waste.
March uses a Trie structure to automatically detect and reuse identical token prefixes. Instead of storing [system_prompt + history] 1000 times, it's stored once. Everyone shares it.
- 80-97% memory reduction in prefix-heavy workloads (tested on SmolLM2-135M with 500 multi-turn conversations)
- Zero-copy queries — returns direct pointers into the memory pool, no expensive memcpy on the hot path
- Predictable memory usage — fixed-size page pool with O(L) complexity
- Trade-off: slightly slower than dict O(1) lookup, but the memory savings are worth it in production

1 reply

·

reacted to reaperdoesntknow's post with 👍 7 months ago

Post

3,309

We present a methodology for training small language models on CPU at FP32 precision
that achieves capability-per-dollar efficiency orders of magnitude beyond GPU-based training.
Across15modelsspanningfournovelarchitecturefamilies—MixtureofAttentions(MoA),cross-
architecture fusion (Qemma), swarm intelligence (SAGI), and metric-space causal language
models (DiscoverLM)—total compute cost was $24 on a single AMD EPYC 9454P proces-
sor. We introduce seven methodological pillars: (1) FP32 precision preservation, with exper-
iments demonstrating 5,810×single-operation error and 23,225×compounding error ratio for
FP16 at network depth; (2) sparse cognitive architectures where 0.02–7% of parameters activate
per token, matching CPU branching rather than GPU SIMD; (3) developmental curriculum
training progressing from language to logic to transfer to depth; (4) continuous belt-fed data
ingestion eliminating truncation waste; (5) hardware-native optimization for AMD Zen 4 via
AOCL/OpenMP/NUMA-aware allocation; (6) self-regulating thermodynamic governance with
emergent temperature measurement grounded in L2-star discrepancy; and (7) open-standard
compute (AVX2 SIMD at FP32) free of proprietary vendor dependency. We argue that transformers were designed for GPU hardware rather than mathematical optimality, and that architecture designed for geometric correctness—metric-space attention, triangle inequality enforcement, sparse expert routing—naturally favor CPU execution. For sub-2B parameter models, CPU training produces more capable models at a fraction of the cost.

6 replies

·

reacted to danielhanchen's post with 🔥 8 months ago

Post

3,982

We collaborated with NVIDIA to teach you about Reinforcement Learning and RL environments. 💚 Learn:

• Why RL environments matter + how to build them
• When RL is better than SFT
• GRPO and RL best practices
• How verifiable rewards and RLVR work

Blog: https://unsloth.ai/blog/rl-environments

4 replies

·

replied to their post 8 months ago

This comment has been hidden

replied to their post 8 months ago

This comment has been hidden

replied to their post 8 months ago

What an academic tone! New baselines, here.

reacted to their post with 🤗 8 months ago

Post

5,424

Arcade-3B — SmolReasoner
NoesisLab/Arcade-3B
Arcade-3B is a 3B instruction-following and reasoning model built on SmolLM3-3B. It is the public release from the ARCADE project at NoesisLab, which investigates the State–Constraint Orthogonality Hypothesis: standard Transformer hidden states conflate factual content and reasoning structure in the same subspace, and explicitly decoupling them improves generalization.

5 replies

·

posted an update 8 months ago

Post

5,424

Arcade-3B — SmolReasoner
NoesisLab/Arcade-3B
Arcade-3B is a 3B instruction-following and reasoning model built on SmolLM3-3B. It is the public release from the ARCADE project at NoesisLab, which investigates the State–Constraint Orthogonality Hypothesis: standard Transformer hidden states conflate factual content and reasoning structure in the same subspace, and explicitly decoupling them improves generalization.

5 replies

·

reacted to smirki's post with 👍 8 months ago

Post

763

Introducing OmniCoder-9B

We trained a 9B coding agent on 425K real agentic trajectories from Claude Opus 4.6, GPT-5.4, GPT-5.3-Codex, and Gemini 3.1 Pro across Claude Code, OpenCode, Codex, and Droid scaffolding.

Results:
- GPQA Diamond: 83.8 pass@1 (166/198), 86.4 pass@3 — above GPT-OSS-120B (80.1), Qwen3.5-9B (81.7), and Claude Haiku 4.5 (73)
- AIME 2025: 90 pass@5 (27/30)
- Terminal-Bench 2.0: 28.1 (25/89) — +8.1 points over base model

The key insight: We trained on what frontier agents actually do, real tool calls, real error recovery, real edit diffs. The model learns read-before-write patterns, responds to LSP diagnostic, and applies minimal diffs instead of full rewrites.

Base: Qwen3.5-9B. LoRA SFT, 4x H200, Axolotl, 99.35% packing efficiency.

Weights:

Tesslate huggingface.co/Tesslate/OmniCoder-9B
GGUF: huggingface.co/Tesslate/OmniCoder-9B-GGUF
Apache 2.0. Run it locally.

reacted to marksverdhei's post with 🤯 8 months ago

Post

762

The hidden gem of open-source embedding models: LCO-Embedding
for text, image AND audio!

I found this model after reading the recent Massive Audio Embedding Benchmark (MAEB) paper, as it blew the other models out of the water on day zero. I've been using it personally for about a week, and searching my files by describing music, sound effects or images is both practical and entertaining. Really underrated model, would highly recommend checking it out: LCO-Embedding/LCO-Embedding-Omni-7B

PS: If you're looking you run this model on llama.cpp, i've gone ahead and quantized them for you here 👉 https://huggingface.co/collections/marksverdhei/lco-embedding-omni-gguf

7 replies

·

reacted to their post with 🤗 8 months ago

Post

1,980

We deleted the Embedding Layer -- INTRO Our Collins-Embedding-3M
NoesisLab/Collins-Embedding-3M
Most "small" models are just giant vocab tables in a trench coat. Collins-3M changes that. By using 2-Universal Hashing and Chernoff-bound noise suppression, we’ve collapsed the embedding space into a fixed O(1) hash-map.
* STSB: 0.7114 (Beating many 100M+ models)
* Size: 3M (Edge-ready, IoT-ready)
* Tech: Randomized Sign-Hashing + RoPE positional injection.
Built by NoesisLab

posted an update 8 months ago

Post

1,980

We deleted the Embedding Layer -- INTRO Our Collins-Embedding-3M
NoesisLab/Collins-Embedding-3M
Most "small" models are just giant vocab tables in a trench coat. Collins-3M changes that. By using 2-Universal Hashing and Chernoff-bound noise suppression, we’ve collapsed the embedding space into a fixed O(1) hash-map.
* STSB: 0.7114 (Beating many 100M+ models)
* Size: 3M (Edge-ready, IoT-ready)
* Tech: Randomized Sign-Hashing + RoPE positional injection.
Built by NoesisLab

reacted to SeaWolf-AI's post with 🔥 8 months ago

Post

5,159

ALL Bench — Global AI Model Unified Leaderboard

FINAL-Bench/all-bench-leaderboard

If you've ever tried to compare GPT-5.2 and Claude Opus 4.6 side by side, you've probably hit the same wall: the official Hugging Face leaderboard only tracks open-source models, so the most widely used AI systems simply aren't there. ALL Bench fixes that by bringing closed-source models, open-weight models, and — uniquely — all four teams under South Korea's national sovereign AI program into a single leaderboard. Thirty-one frontier models, one consistent scoring scale.
Scoring works differently here too. Most leaderboards skip benchmarks a model hasn't submitted, which lets models game their ranking by withholding results. ALL Bench treats every missing entry as zero and divides by ten, so there's no advantage in hiding your weak spots.
The ten core benchmarks span reasoning (GPQA Diamond, AIME 2025, HLE, ARC-AGI-2), coding (SWE-bench Verified, LiveCodeBench), and instruction-following (IFEval, BFCL). The standout is FINAL Bench — the world's only benchmark measuring whether a model can catch and correct its own mistakes. It reached rank five in global dataset popularity on Hugging Face in February 2026 and has been covered by Seoul Shinmun, Asia Economy, IT Chosun, and Behind.
Nine interactive charts let you explore everything from composite score rankings and a full heatmap to an open-vs-closed scatter plot. Operational metrics like context window, output speed, and pricing are included alongside benchmark scores.
All data is sourced from Artificial Analysis Intelligence Index v4.0, arXiv technical reports, Chatbot Arena ELO ratings, and the Korean Ministry of Science and ICT's official evaluation results. Updates monthly.

reacted to abdurrahmanbutler's post with 🤗 8 months ago

Post

2,600

🚀 𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝗞𝗮𝗻𝗼𝗻 𝟮 𝗘𝗻𝗿𝗶𝗰𝗵𝗲𝗿: 𝘁𝗵𝗲 𝘄𝗼𝗿𝗹𝗱’𝘀 𝗳𝗶𝗿𝘀𝘁 𝗵𝗶𝗲𝗿𝗮𝗿𝗰𝗵𝗶𝗰𝗮𝗹 𝗴𝗿𝗮𝗽𝗵𝗶𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹

Today we’re publicly releasing Kanon 2 Enricher, and with it, an entirely new class of AI model that we’re calling a hierarchical graphitization model.
This is fundamentally different from both universal extraction models and generative models.

As a hierarchical graphitization model, Kanon 2 Enricher natively outputs a 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗴𝗿𝗮𝗽𝗵 rather than tokens, which makes it architecturally incapable of hallucinating or inventing text that wasn’t present in the input.

What that enables in practice is unlike any other model or ML architecture on the market:

• 𝗡𝗼 𝗵𝗮𝗹𝗹𝘂𝗰𝗶𝗻𝗮𝘁𝗶𝗼𝗻𝘀 🤖
It cannot hallucinate. All references and links are stored as spans, meaning exact character offsets anchored to the original text.

• 𝗛𝗶𝗲𝗿𝗮𝗿𝗰𝗵𝗶𝗰𝗮𝗹 𝘀𝗲𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻, 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝗲𝘅𝘁𝗿𝗮𝗰𝘁𝗶𝗼𝗻 📑
It deconstructs a document’s full nested hierarchy, down to chapters, sections, clauses, schedules, signatures, and even singular sentences, and classifies each span with dozens of contextual features.

• 𝗘𝗻𝘁𝗶𝘁𝘆 𝗲𝘅𝘁𝗿𝗮𝗰𝘁𝗶𝗼𝗻, 𝗱𝗶𝘀𝗮𝗺𝗯𝗶𝗴𝘂𝗮𝘁𝗶𝗼𝗻, 𝗮𝗻𝗱 𝗹𝗶𝗻𝗸𝗶𝗻𝗴 🔗
It resolves what references actually point to, then links entities, citations, and cross-references into a single coherent graph.

• 𝗚𝗿𝗮𝗽𝗵-𝗳𝗶𝗿𝘀𝘁 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆 🏃‍➡️
Small enough to run locally on a consumer PC with sub-second latency, and it stays reliable on long documents where front

To read more about our new model, check out our latest Hugging Face article:
https://huggingface.co/blog/isaacus/introducing-kanon-2-enricher

reacted to their post with 🤗 8 months ago

Post

4,836

🔥 UPGRADE in Kai: 30B Scaling! 🔥
NoesisLab/Kai-30B-Instruct
NoesisLab/Kai-30B-Instruct
We are incredibly excited to announce that the Kai-30B-Instruct model and its official Space are now LIVE! 🚀
If you've been following the journey from Kai-0.35B to Kai-3B, you know we're rethinking how models reason. Tired of verbose, slow Chain-of-Thought (CoT) outputs that flood your screen with self-talk? So are we.
Kai-30B-Instruct scales up our Adaptive Dual-Search Distillation (ADS) framework. By bridging classical A* heuristic search with continuous gradient descent , we use an information-theoretic log-barrier to physically prune high-entropy reasoning paths during training.
The result? Pure implicit reasoning. The model executes structured logic, arithmetic carries, and branch selections as a reflex in a single forward pass—no external scaffolding required.
At 3B, we observed a phase transition where the model achieved "logical crystallization". Now, at 30B, we are giving the ADS regularizer the massive representational capacity it needs to tackle higher-order symbolic abstractions and complex reasoning tasks.
🧪 Test Kai yourself in our new Space:
NoesisLab/Kai-30B-Instruct
📦 Model Weights:
NoesisLab/Kai-30B-Instruct
Bring your hardest math, logic, and coding benchmarks. We invite the community to stress-test the limits of the penalty wall! 🧱💥

1 reply

·

Zixi "Oz" Li

AI & ML interests

Recent Activity

Organizations

Zixi "Oz" Li

AI & ML interests

Recent Activity

Organizations

OzTianlu's activity