Avi Fenesh

Avifenesh

https://inference.tiyuvta.ai/

AI & ML interests

Inference and quant research, fine-tuning and specialist models, BW, self-hosting, on-prem, and intelligence open for everyone

Recent Activity

updated a model about 13 hours ago

tiyuvta/memra-bench

updated a model about 13 hours ago

tiyuvta/gemma-4-31B-it-assistant-MTP-GGUF

updated a model about 13 hours ago

tiyuvta/Qwen3.8-Flash-Next-NVFP4

View all activity

Organizations

New activity in tiyuvta/Ornith-1.5-35B-A3B-NVFP4-MTP-GGUF 8 days ago

Any chance of uploading Ornith1.5-9B-NVFP4-MTP-GGUF

#1 opened 22 days ago by

sa13ma

New activity in tiyuvta/GLM-5.3-Flash-NVFP4 13 days ago

Context section correction: 1M demonstrated, chunk-schedule fix, honest boundaries

#4 opened 13 days ago by

Avifenesh

Context section correction

#3 opened 13 days ago by

Avifenesh

Context section correction

#2 opened 13 days ago by

Avifenesh

Context section: 1,035,357 tokens demonstrated (1m-demo receipts); the monolithic-prefill wall is fixed by the chunk schedule; honest 4-card/latency boundaries stated

#1 opened 13 days ago by

Avifenesh

New activity in tiyuvta/Ornith-1.5-35B-A3B-NVFP4-MTP-GGUF 16 days ago

docs: attribute the hosted Ornith link

#3 opened 16 days ago by

Avifenesh

docs: link the hosted Ornith API

#2 opened 16 days ago by

Avifenesh

New activity in tiyuvta/Qwen3.8-27B-NVFP4-MTP-GGUF 16 days ago

docs: link the hosted Qwen3.8 API

#3 opened 16 days ago by

Avifenesh

New activity in ornith-ai/Ornith-1.5-35B-A3B 20 days ago

No speculative decoding, No QAT

#21 opened 21 days ago by

engrtipusultan

New activity in google/gemma-4-31B-it 23 days ago

When Gemma doesn't have the right tool schema it hallucinates it, create outputs of the fake-calls and use them as memory (WIP)

👀👍 4

#136 opened about 2 months ago by

luislozanog86

New activity in unsloth/Qwen3.8-27B-GGUF 24 days ago

DFlash2?

👀 5

#79 opened 24 days ago by

guarism0

NVFP4/MXFP4 GGUF Version?

👍 3

#75 opened 24 days ago by

CYISNOTHERE

New activity in ornith-ai/Ornith-1.5-35B-A3B 24 days ago

Unsloth GGUF's please

#3 opened 24 days ago by

jlindeberg

New activity in ornith-ai/Ornith-1.5-35B-A3B-GGUF 24 days ago

MTP Support

🚀 9

#2 opened 24 days ago by

DefaultDF

New activity in Qwen/Qwen3.8-27B 25 days ago

MTP (SPEC_DRAFT_N_MAX) leads to Nvidia RC Watchdog killing the kernel, llama.cpp calls SIGABRT

➕ 1

#122 opened 27 days ago by

Nosepicker3

New activity in google/gemma-4-31B-it 25 days ago

Gemma-4-31B-it occasionally emits <turn|> token when served via vLLM, causing failures when using it with GitHub Copilot Custom Model Provider

#138 opened about 2 months ago by

aygupt

New activity in unsloth/Qwen3.8-27B-GGUF 27 days ago

Small Quants with integrated MTP dont make sense

👍 5

#37 opened 29 days ago by

TheWegemann

New activity in Qwen/Qwen3.8-27B 27 days ago

qwen3.8-27b on one pro 6000: 140 tok/s, day after release

#101 opened 28 days ago by

Avifenesh

MTP speculation is net-negative under Ollama's GGUF path, strongly positive under MLX (measured on M4 Pro)

#80 opened 29 days ago by

OtsoaD

New activity in tiyuvta/Qwen3.8-27B-NVFP4-MTP-GGUF 28 days ago

live api

#2 opened 28 days ago by

Avifenesh

Avi Fenesh

AI & ML interests

Recent Activity

Organizations

Avifenesh's activity

Any chance of uploading Ornith1.5-9B-NVFP4-MTP-GGUF

Context section correction: 1M demonstrated, chunk-schedule fix, honest boundaries

Context section correction

Context section correction

Context section: 1,035,357 tokens demonstrated (1m-demo receipts); the monolithic-prefill wall is fixed by the chunk schedule; honest 4-card/latency boundaries stated

docs: attribute the hosted Ornith link

docs: link the hosted Ornith API

docs: link the hosted Qwen3.8 API

No speculative decoding, No QAT

When Gemma doesn't have the right tool schema it hallucinates it, create outputs of the fake-calls and use them as memory (WIP)

DFlash2?

NVFP4/MXFP4 GGUF Version?

Unsloth GGUF's please

MTP Support

MTP (SPEC_DRAFT_N_MAX) leads to Nvidia RC Watchdog killing the kernel, llama.cpp calls SIGABRT

Gemma-4-31B-it occasionally emits <turn|> token when served via vLLM, causing failures when using it with GitHub Copilot Custom Model Provider

Small Quants with integrated MTP dont make sense

qwen3.8-27b on one pro 6000: 140 tok/s, day after release

MTP speculation is net-negative under Ollama's GGUF path, strongly positive under MLX (measured on M4 Pro)

live api