Hugging Face's logo
Activity Feed

AI & ML interests

None defined yet.

Recent Activity
honza-rosecky  updated a model 6 days ago
thinletter/qwen3-embedding-0.6b-vq-clients
honza-rosecky  updated a model 6 days ago
thinletter/harrier-0.6b-vq-clients
honza-rosecky  updated a Space 9 days ago
thinletter/demo
View all activity
honza-rosecky 
updated 2 models 6 days ago
honza-rosecky 
updated 2 Spaces 9 days ago
honza-rosecky 
in thinletter/harrier-0.6b-vq-clients 11 days ago

Prompt-K/V variants added: four files calibrated with the query prompt's keys and values in full precision
#5 opened 11 days ago by
honza-rosecky
honza-rosecky 
in thinletter/qwen3-embedding-0.6b-vq-clients 11 days ago

First Czech vector-quantised client: 237 MiB, 96.8 % of fp32 nDCG@10 on WebFAQ-cs, runs in the browser
#1 opened 11 days ago by
honza-rosecky
honza-rosecky 
published a model 11 days ago
honza-rosecky 
published a Space 11 days ago
honza-rosecky 
updated 3 models 11 days ago
honza-rosecky 
in thinletter/harrier-0.6b-vq-clients 11 days ago

The runtime is now open source: quantiser, .vqw compiler and WebGPU runtime (Apache-2.0)
#4 opened 11 days ago by
honza-rosecky
honza-rosecky 
in thinletter/harrier-0.6b-vq-clients 16 days ago

The query prompt's K/V in full precision: +0.01–0.04 nDCG@10 at 1.6–1.8 bits per weight for 2 MB of data (measured, not yet in the files)

2
#3 opened 17 days ago by
honza-rosecky
honza-rosecky 
published a Space 17 days ago
honza-rosecky 
in thinletter/harrier-0.6b-vq-clients 17 days ago

Vector-quantised harrier-0.6b query clients at 1.8–2.1 bits per weight (105–127 MiB) — summary and links
#2 opened 17 days ago by
honza-rosecky
honza-rosecky 
in thinletter/harrier-0.6b-query-clients 17 days ago

harrier-0.6b query clients: 192–235 MiB, 94–99 % of fp32 on the unchanged index — summary and links
#1 opened 17 days ago by
honza-rosecky
honza-rosecky 
in thinletter/bge-m3-query-clients 17 days ago

bge-m3 query-side clients at 3–4 bits, calibrated on Czech text — summary and links
#2 opened 17 days ago by
honza-rosecky
honza-rosecky 
in thinletter/qwen3-embedding-0.6b-query-clients 17 days ago

Query-side clients for Qwen3-Embedding-0.6B: keep your index, shrink the query encoder — summary and links
#3 opened 17 days ago by
honza-rosecky
honza-rosecky 
in thinletter/llm-weight-compression-evidence 17 days ago

Which objective should a post-training quantizer optimize? — summary and links
#2 opened 17 days ago by
honza-rosecky
honza-rosecky 
in thinletter/llm-weight-compression-explorer 17 days ago

links: moved to thinletter

1
#1 opened 17 days ago by
honza-rosecky