Repository navigation

All pull requests

Labels (9)Milestones (0)New pull request

Open

(32)

Closed

(244)

Run the prefill shared expert beside the MoE routing
#443zaiyang-zhang opened Oct 16, 2026Contributor
Run the engine on Apple M1 and M2 within their driver's thread caps
#442victorgabr opened Oct 16, 2026
Key the backend regression's short prefills by their prompt length
#439jianc99 opened Oct 15, 2026Collaborator
Pass a decode cycle's verify rows as a parameter
#438jianc99 opened Oct 15, 2026Collaborator
Run the dense prefill FFN in int8 on the GPU of an M5 or later
#436liang2kl opened Oct 15, 2026Contributor
Share the full chunk's attention budget with short prefill chunks
#434adriangalilea opened Oct 14, 2026
Stretch the context window with YaRN (--yarn-factor)
#431jianc99 opened Oct 14, 2026Collaborator
Retain bounded optional RAM checkpoints for historical prefix reuse
#430ardnt opened Oct 14, 2026
Draft Flash-Next's MTP proposals from a draft vocabulary and the request's context
#429Vardaan97 opened Oct 14, 2026
2
Report thinking tokens and timings in Messages and Responses
#427zaiyang-zhang opened Oct 14, 2026Contributor
Serve custom (freeform) Responses tools
#426zaiyang-zhang opened Oct 14, 2026Contributor
Relay requests for other models to an upstream API
#425zaiyang-zhang opened Oct 14, 2026Contributor
Load ahead in Flash-Next's hyper-connection norm
#423Vardaan97 opened Oct 14, 2026
2
Stream Qwen3.8-Flash-Next's PLE table from its GGUF with --ple-ssd-cache
#410Vardaan97 opened Oct 13, 2026
2
Check the remembered Neural Engine calibration at each start, and calibrate faster
#407liang2kl opened Oct 13, 2026Contributor
Bound prefill while another Splash server on the Mac decodes
#377jianc99 opened Oct 11, 2026Collaborator
Run decode's MoE expert passes on their tiles' live rows on Apple9
#367janhilgard opened Oct 11, 2026Contributor
1
Add --kv-format zbf16: lossless BF16 KV in 0.81 of the memory
#340coco-alen opened Oct 7, 2026
Reuse local models and support local GGUF vision projectors
#338yannip1234 opened Oct 7, 2026Contributor
Add prefill-only decision scoring: POST /v1/score and /v1/decisions
#295janhilgard opened Oct 4, 2026Contributor
3
Skip expert work for low-confidence greedy draft suffixes
#283linson007 opened Oct 3, 2026Contributor
3
Add Qwen3.8-Flash-Next (qwen4exp) GGUF support with MTP drafting
#253janhilgard opened Oct 2, 2026Contributor
56
research: preserve PQ2 M8 target-forward candidates and paired results
#252zaiyang-zhang opened Oct 2, 2026Contributor
Tune GGUF decode splits for Apple11 (M6) by format
#235zaiyang-zhang opened Sep 30, 2026Contributor
Add Oh My Pi coding agent support
#230yannip1234 opened Sep 29, 2026Contributor

Pull requests