The 80B Qwen3-Coder-Next on a 16 GB Mac, about 13 tok/s, fully offline. A mixture-of-experts inference engine for Apple silicon.
Rust
Commits on main
0.5.1 docs: MODELS.md coder section leads with the squeezed file
1fd75b5 1picassoai committed 1 month ago
0.5.1 docs: the tester's numbers (about 12 tok/s, 11.5 at 13K), disk numbers, split marked not yet measured
37f2ac1 1picassoai committed 1 month ago
0.5.1 docs: the ready-squeezed 80B coder on the front page, release notes with the measured numbers (HF link to fill)
698396e 1picassoai committed 1 month ago
README: the four gpt-oss setup steps on the front page; the numbers say nobody clicks through
6733e67 1picassoai committed 1 month ago
README: the three requirements on the first screen, under the install line
e722c9e 1picassoai committed 1 month ago
README: the promise, the install line and the demo on the first screen
6b01159 1picassoai committed 1 month ago
docs: demo GIF under a new name so cached copies of the old one are not served
a56ab9c 1picassoai committed 1 month ago
docs: README demo GIF is gpt-oss-120b on the 0.5.0 release build
fa42a2f 1picassoai committed 1 month ago
docs: front page is what, why and install; every model on the models page; gpt-oss page as a wiki page
94e0d02 1picassoai committed 1 month ago
docs/GPT-OSS.md: the task section itself (the previous commit carried only the code file)
faebb7d 1picassoai committed 1 month ago
docs/GPT-OSS.md: the task, the code it wrote at that speed, and the 20 checks
24a1d6a 1picassoai committed 1 month ago
Stand-in gap (--stand-ins as shipped), the numbers from three runs, --max-tokens 8000 in the gpt-oss commands
ffdd2b2 1picassoai committed 1 month ago
--stand-ins guesses one expert a layer for gpt-oss: 1,000-word answers 8.4 -> 10.9 tok/s
4569d03 1picassoai committed 1 month ago
docs: the gpt-oss squeeze takes about 13 minutes on an M4 (776 s measured, output byte-identical to the tested file)
07869d9 1picassoai committed 1 month ago
0.5.0: gpt-oss-120b - the family, its squeeze, --stand-ins and --flash-reads
64e4ff6 1picassoai committed 1 month ago
tools/package-release.sh: builds the ready-built download with no build paths inside
cd343b5 1picassoai committed 1 month ago
README: badges that apply (release, license, macOS on Apple silicon)
4a442b5 1picassoai committed 1 month ago
README: after the squeeze, delete the download (Coachwhip only uses the squeezed file); disk need stated as 105 GB while squeezing, 46 GB after
fae7fc1 Merlin committed 1 month ago
README: the 122B first in the models table, What you need and Bring a model; no model is labelled recommended
5d641a0 Merlin committed 1 month ago
MODELS.md leads with the 122B: its section moves to the top, word for word
8ee0037 Merlin committed 1 month ago
README leads with the 122B and its table; demo GIF is the 122B in real time; docs: no promises past the frozen range, no stale or placeholder lines
c4e0456 Merlin committed 1 month ago
Merge public main (be00834, the installer pinned to the signed tag) into release/0.4.0; pin moves to v0.4.0, README opens with 0.4.0
96bb3b0 Merlin committed 1 month ago
docs: notes state what 0.4.0 is measured on (122B, 80B) and what it is not (35B, beyond 12K, exact 6K, disk-full leftover)
6db0d4e Merlin committed 1 month ago
docs: the agent-task result as a note (fast 5 of 6, exact passes the sixth), per the tester
7f89b19 Merlin committed 1 month ago