Dwarf Star Support for Qwen3.8 Flash Next Developer ivanfioravanti merged a pull request adding Qwen3.8 Flash Next support to the antirez/ds4 repository, including qwen4exp loading, tokenizer and chat formatting, PLE sidecars, Metal text and vision graphs, built-in MTP, and recurrent KV persistence. The branch was force-pushed multiple times between September 6 and September 13, 2026, and a contributor flagged that the Q4_1 encoder from commit a0b1658 was dropped in b020c9f and the squashed 3030554 while qwen4_pack.py still uses it for the PLE sidecar. A Q4_K imatrix test model was published at huggingface.co/ivanfioravanti/Qwen3.8-Flash-Next-DS4-Q4, and the PR includes reproducible throughput charts and measurements on M3 Ultra. Qwen3.8 flash next - 991 Merged Merged Conversation Contributor Author | Q4 K imatrix model for testing is here https://huggingface.co/ivanfioravanti/Qwen3.8-Flash-Next-DS4-Q4 | ivanfioravanti https://github.com/ivanfioravanti force-pushed https://github.com/antirez/ds4/compare/ecf1da3c9516d0d78a31c61ba8cba8b535488caa..236cb2a549d05ea1941a19bd4f154e0abc858661 the qwen3.8-flash-next branch from to https://github.com/antirez/ds4/commit/ecf1da3c9516d0d78a31c61ba8cba8b535488caa ecf1da3 236cb2a Compare https://github.com/antirez/ds4/compare/ecf1da3c9516d0d78a31c61ba8cba8b535488caa..236cb2a549d05ea1941a19bd4f154e0abc858661 September 6, 2026 10:33 event-30637686565 This was referenced Sep 6, 2026 evandhoffman https://github.com/evandhoffman added a commit to evanwtf/local-llm that referenced this pull request Sep 6, 2026 ref-commit-cd82bc2 ivanfioravanti https://github.com/ivanfioravanti marked this pull request as draft September 6, 2026 22:22 event-30655094735 Contributor Author | Finalizing Vision, Q2 and tools after rebase. | ivanfioravanti https://github.com/ivanfioravanti force-pushed https://github.com/antirez/ds4/compare/8f7c4fa65457c1f7d097a8fd5882e8e30c41575e..ffd85d426313ace6dae805e9f2fb4424d5b427fd the qwen3.8-flash-next branch from to https://github.com/antirez/ds4/commit/8f7c4fa65457c1f7d097a8fd5882e8e30c41575e 8f7c4fa ffd85d4 Compare https://github.com/antirez/ds4/compare/8f7c4fa65457c1f7d097a8fd5882e8e30c41575e..ffd85d426313ace6dae805e9f2fb4424d5b427fd September 6, 2026 23:11 event-30656157149 ivanfioravanti https://github.com/ivanfioravanti marked this pull request as ready for review September 6, 2026 23:18 event-30656299619 evandhoffman https://github.com/evandhoffman added a commit to evanwtf/local-llm that referenced this pull request Sep 7, 2026 ref-commit-4ce9105 evandhoffman https://github.com/evandhoffman added a commit to evanwtf/local-llm that referenced this pull request Sep 8, 2026 ref-commit-4537b83 This was referenced Sep 11, 2026 ivanfioravanti https://github.com/ivanfioravanti force-pushed https://github.com/antirez/ds4/compare/ff4f0ff4fdff70d6b7c3941ef437b91dde960e14..30305544f96cc35357e0dd376eeff2b1ab3e365e the qwen3.8-flash-next branch from to https://github.com/antirez/ds4/commit/ff4f0ff4fdff70d6b7c3941ef437b91dde960e14 ff4f0ff 3030554 Compare https://github.com/antirez/ds4/compare/ff4f0ff4fdff70d6b7c3941ef437b91dde960e14..30305544f96cc35357e0dd376eeff2b1ab3e365e September 12, 2026 23:32 event-31036732440 | @ivanfioravanti https://github.com/ivanfioravanti Quick heads-up: the Q4 1 encoder from a0b1658 appears to have been dropped in b020c9f and the squashed 3030554 , while qwen4 pack.py still uses it for the PLE sidecar. The current native quantizer marks Q4 1 unsupported. The original implementation likely just needs to be restored before upstreaming. | Contributor Author | You rock Gonna fix this | ivanfioravanti https://github.com/ivanfioravanti force-pushed https://github.com/antirez/ds4/compare/d8ba365815af259dc363f991cc1a47dff7637ccb..cf5f223d14ca434142cfbeaaf7eeb9353fee10ba the qwen3.8-flash-next branch 3 times, most recently from to https://github.com/antirez/ds4/commit/d8ba365815af259dc363f991cc1a47dff7637ccb d8ba365 cf5f223 Compare https://github.com/antirez/ds4/compare/d8ba365815af259dc363f991cc1a47dff7637ccb..cf5f223d14ca434142cfbeaaf7eeb9353fee10ba September 13, 2026 09:21 event-31048305644 Implement qwen4exp loading, tokenizer and chat formatting, PLE sidecars, Metal text and vision graphs, built-in MTP, and recurrent KV persistence. Integrate the CLI, agent, server, download targets, and conversion tools. Preserve the native Q4 1 encoder, Q4 K accumulation guard, and shared quantizer build target. Keep exact-MTP rewind and vision-cache row-width fixes. Persist valid logits and recurrent state at completed prefill chunks, and reuse Qwen text-answer history when tools clients omit reasoning. Retain exact reasoning replay and test continued checkpoints, cancellation, cache eviction, and restart reuse with and without reasoning content. Include reproducible charts and measurements comparing throughput before and after the checkpoint fixes on M3 Ultra. Add Qwen trunk steering and prompt activation capture, including GDN layers. Keep the MTP predictor outside the 48-row bank and preserve prompt dumps through one-token prefill tails and multi-row speculative verification. Recover Chrome target creation and retain its last page target. Preserve full agent transcripts in stripped-session format when KV no longer fits, while retaining the newer automatic compaction and continuation logic. Add steering/MTP/capture, Chrome CDP recovery, and full-session save/reload regressions. Adapt the intended changes from fork PRs 13 and 7. Reconcile with antirez main bd66c40, preserving DeepSeek v4.1 model IDs, tool syntax, session state, and Metal kernel dispatch. Exclude superseded engine code and prototypes. Co-authored-by: Andreas Parusel