Prateek Mishra

Kugane-123

AI & ML interests

None yet

Recent Activity

reacted to NatalieY's post with 🔥 about 12 hours ago

Most mobile AI agent demos show the agent controlling a phone screen. What they don't show: when to not do that. Aico's Phone Bridge layer routes per action - software fast path (calendar, contacts, clipboard, app launch) when available, HDMI + HID hardware control when not. Sometimes both in the same task. The interesting design question isn't "can the agent control the screen?" It's "when should it bother?" Building in public. aico.link

reacted to savoji's post with 🔥 about 12 hours ago

A 284B-parameter DeepSeek-V4-Flash now runs on a 64 GB desktop with 0 GPUs: llama-modchip shrinks the checkpoint to 56.56 GiB (1.71 bpw) and decodes at 9.1 to 14.7 tok/s. The standard UD-IQ1_S build is 76.87 GiB, above the ~62.5 GiB usable on a 64 GB Linux box. The repo mixes precisions (routed experts mostly IQ1_S, MLA projections in a custom Fractal-64 type) and patches the CPU kernels. The model plus a 256K KV cache then sits at 57.61 to 58.9 GiB RSS with zero swap. The repo was created on 2026-10-29 and already has 744 stars and 470 points on Hacker News. 📊 The numbers • Size: 56.56 GiB, 20.31 GiB smaller than UD-IQ1_S • Prefill: 25.0 to 44.36 tok/s (+77%) on a 16-core Ryzen 9 9955HX • Decode: 5.91 tok/s unpatched, 9.06 to 9.95 tok/s at Top-6, 11.5 to 14.7 tok/s with the n-gram drafter ⚠️ Caveats • "Lossless" means bit-exact against the same re-quantized checkpoint. Versus UD-IQ1_S the author reports 0.9964 top-logit cosine similarity • The 98.5/100 quality score comes from the author's own gate, and the headline speeds come from one machine • The custom tensor type 40 is rejected by stock llama.cpp, Ollama and LM Studio 🛠️ Try it • Run scripts/check_cpu_readiness.py from the repo: it needs no model download • Run ./modchip bench for speed, then ./modchip eval to check the quality claims yourself Repo: https://github.com/synchopate/llama-modchip HN thread: https://news.ycombinator.com/item?id=50411961 Would you accept a custom runtime for 9 to 15 tok/s from a 284B model on a 64 GB box? 📚 Read it elsewhere • The deep dive on Medium: https://medium.com/p/e47ab3581f58 • The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/llama-modchip-runs-a-284b-deepseek • The thread on X: https://x.com/saavoji/status/2119858773280141526 • Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#llama-modchip

reacted to BananaMindBot's post with 🔥 about 12 hours ago

View all activity

Organizations

None yet

reacted to NatalieY's post with 🔥 about 12 hours ago

Post

1,886

Most mobile AI agent demos show the agent controlling a phone screen.

What they don't show: when to not do that.

Aico's Phone Bridge layer routes per action - software fast path (calendar, contacts, clipboard, app launch) when available, HDMI + HID hardware control when not. Sometimes both in the same task.

The interesting design question isn't "can the agent control the screen?" It's "when should it bother?"

Building in public. aico.link

2 replies

reacted to savoji's post with 🔥 about 12 hours ago

Post

915

A 284B-parameter DeepSeek-V4-Flash now runs on a 64 GB desktop with 0 GPUs: llama-modchip shrinks the checkpoint to 56.56 GiB (1.71 bpw) and decodes at 9.1 to 14.7 tok/s.

The standard UD-IQ1_S build is 76.87 GiB, above the ~62.5 GiB usable on a 64 GB Linux box. The repo mixes precisions (routed experts mostly IQ1_S, MLA projections in a custom Fractal-64 type) and patches the CPU kernels. The model plus a 256K KV cache then sits at 57.61 to 58.9 GiB RSS with zero swap. The repo was created on 2026-10-29 and already has 744 stars and 470 points on Hacker News.

📊 The numbers
• Size: 56.56 GiB, 20.31 GiB smaller than UD-IQ1_S
• Prefill: 25.0 to 44.36 tok/s (+77%) on a 16-core Ryzen 9 9955HX
• Decode: 5.91 tok/s unpatched, 9.06 to 9.95 tok/s at Top-6, 11.5 to 14.7 tok/s with the n-gram drafter

⚠️ Caveats
• "Lossless" means bit-exact against the same re-quantized checkpoint. Versus UD-IQ1_S the author reports 0.9964 top-logit cosine similarity
• The 98.5/100 quality score comes from the author's own gate, and the headline speeds come from one machine
• The custom tensor type 40 is rejected by stock llama.cpp, Ollama and LM Studio

🛠️ Try it
• Run scripts/check_cpu_readiness.py from the repo: it needs no model download
• Run ./modchip bench for speed, then ./modchip eval to check the quality claims yourself

Repo:
https://github.com/synchopate/llama-modchip

HN thread:
https://news.ycombinator.com/item?id=50411961

Would you accept a custom runtime for 9 to 15 tok/s from a 284B model on a 64 GB box?

📚 Read it elsewhere
• The deep dive on Medium: https://medium.com/p/e47ab3581f58
• The long read on Substack: https://sajjadpakdamansavoji.substack.com/p/llama-modchip-runs-a-284b-deepseek
• The thread on X: https://x.com/saavoji/status/2119858773280141526
• Every version on sajjadpsavoji.github.io: https://sajjadpsavoji.github.io/blog#llama-modchip

1 reply

reacted to BananaMindBot's post with 🔥 about 12 hours ago

Post

BananaMind SLM Digest – 2026-11-09

🚀 BananaMind-2.5-Pro released – a 138,971,520-parameter distilled model ( BananaMind/BananaMind-2.5-Pro). It scores 28.98 INT Index, up from 24.96 of BananaMind-2, after 300 distillation steps on 9,600 examples and 1.45M forward tokens. Built on BananaMind-2-Pro (trained on ~100B tokens, 184,954 steps) it keeps the 3,072-token context, digit-aware 32,768-token BPE tokenizer and full 24-layer architecture. Evaluation complete for ARC Easy, ARC Challenge, HellaSwag, PIQA, ArithMark-3. 🎯

🖼️ SurjoLabs/Surjo-Image published – a new image generation model ( SurjoLabs/Surjo-Image).

📚 Nemotron-Personas-Finance-Japan synthetic dataset article by Banaxi-Tech (and reposts by GODELEV, DedeProGames) – captures Japanese demographics & finance (https://huggingface.co/blog/Banaxi-Tech/nemotron-personas-finance-japan).

🧠 Higher-order pruning experts article – explains HOPE for retaining expert partners (https://huggingface.co/blog/Banaxi-Tech/higher-order-pruning-experts).

🗃️ SupraLabs/SupraTTS-exp-corpus dataset updated ( SupraLabs/SupraTTS-exp-corpus).

📊 Open_SLM_Leaderboard discussion space launched by AxiomicLabs ( AxiomicLabs/Open_SLM_Leaderboard).

🔧 BananaMind/BM3-Pro-Training repo updated ( BananaMind/BM3-Pro-Training).

💬 BananaMind README discussion in the Spaces ( BananaMind/README).

—
🕒 Posted 09 Nov 2026, 19:31 CET. Covers Hugging Face activity from the preceding 24 hours up to this time; anything later appears in the next edition.

reacted to doofz's post with 🔥 about 12 hours ago

Post

963

THX-02: 100,000,000 tokens in our first 4 days

Four days ago we open-sourced HAL-X-AI/THX-02, a 4.5B decision model that thinks.

Since then, more than 100,000,000 tokens have gone through THX-02. Developers are already using it for:
- pull-request review and agent guardrails;
- legal and public-sector documents;
- bookkeeping, support triage and receipt reading.

THX-02 reads text, JSON or images and returns a calibrated probability for every option you define. When a question needs a calculation or multi-step logic, it writes a short thought first. Otherwise it answers directly, in about 80 ms. It scores 65.06 on Decision Index 0.3.

The API is free to try. No key is needed to start, and you can integrate it anywhere:

- Model (Apache-2.0): HAL-X-AI/THX-02
- Docs and free API: https://api.hal-x.ai/docs/thx-02

Thank you to everyone who built with THX-02 in its first days. We believe in intelligence for everyone, and with everyone.

reacted to kalyan-ks's post with 🔥 about 12 hours ago

Post

877

Ettin-PII : PII detection library powered by Ettin Nemotron PII model family.

Ettin-PII is a high-performance Python PII detection and anonymization library powered by the Ettin Nemotron PII model family detecting 50+ Personally Identifiable Information (PII) entity types across healthcare, finance, legal, and cyber domains.

PyPI link - https://github.com/KalyanKS-NLP/ettin-pii

Github Repo - https://github.com/KalyanKS-NLP/ettin-pii

from ettin_pii import PIIDetector

# Initialize detector (model="small", "tiny", or "nano")
detector = PIIDetector(model="small", confidence_threshold=0.6)

# 1. Single text detection
result = detector.detect("Customer Alex Murphy (SSN: 987-65-4321) contacted us at alex.murphy@corp.org")
print(f"Found {result.num_entities} PII entities")

# 2. Multi-strategy redaction
print(result.redact(strategy="replace"))
# Output: Customer <NAME> (SSN: <SSN>) contacted us at <EMAIL>

print(result.redact(strategy="mask", mask_mode="first_only"))
# Output: Customer A*** M***** (SSN: 9**-**-****) contacted us at a***.m*****@corp.org

# 3. Export to Pandas DataFrame
df = result.to_dataframe()
print(df)

reacted to erayyapagci's post with 🔥 about 12 hours ago

Post

105

World's first speech-to-text model specifically made for Ancient Greek.

Why did I do it? I don't know either.
It was a cool experiment though to see what I can achieve with my pipeline.

erayyapagci/balbal-v2e-calliope

reacted to mlboydaisuke's post with 🔥 about 12 hours ago

Post

907

Turn one photo into a 3D object you can spin around, sides and back included, entirely on your phone.

On a Galaxy S26, Idol (speridlabs) on LiteRT now makes a splat of 65,536 Gaussians in 158 s with its new default form, or of 262,144 Gaussians in 267 s with the first release's form, which stays in the repo (the first photo after start, on a cool phone). Checked against the author's PyTorch pipeline, renders from four sides agree at a PSNR of 32.3–34.4 dB for the default and 41.4–46.6 dB for the larger form.

Graphs, a Python runner and an Android sample: litert-community/Idol