Hebrus v0.2.0 — Historical DwarfStar release

DwarfStar v0.2.0

Release date: 2026-07-20

This release makes the fork's production path deliberately narrow: local
Apple Metal, at least 64 GB unified memory, and one embedded ExpertMajor v2
artifact for each qualified model family. Canonical GGUFs, sidecars,
ExpertMajor v1, distributed inference, CUDA, and ROCm are not compatibility
paths in this release.

Qualified models

Model	Normal 64 GB startup	Qualified artifact
Qwen3.6-35B-A3B	AUTO selects resident when the pressure gate passes	Qwen3.6-35B-A3B-DS4-ExpertMajor-v2-Q4_K_S.gguf
DeepSeek V4 Flash	AUTO selects resident or SSD from the live memory plan	DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-DS4-ExpertMajor-v2.gguf
GLM 5.2	AUTO always selects the qualified SSD Gold path	GLM-5.2-DS4-ExpertMajor-v2-Q2_K.gguf

Download one exact release artifact and start it without model-admission or
cache-tuning flags:

./download_model.sh qwen-v2
./ds4 -m gguf/Qwen3.6-35B-A3B-DS4-ExpertMajor-v2-Q4_K_S.gguf \
  --ctx 8192 -p "Hello"

Use deepseek-v2 or glm-v2 for the other families. The downloader pins a
versioned model-repository release tag and verifies the complete byte size and
SHA-256 before reporting success. The same normal command shape applies to
ds4-server; residency and the safe expert-cache tier are runtime decisions.

Runtime changes

• Qwen uses the exact scalar Q8 Metal decode path. A final zero-swap 32K
  A/B/B/A preserved byte-identical evidence but found the paired-Q8 experiment
  neutral-to-worse, so its M5-only kernel and test were removed.
• DeepSeek uses a pressure-gated phase policy: 259 experts during prefill,
  4,129 through 32K decode, and 2,065 at 65K and above on the qualified 64 GB
  M5 lane. The old AUTO plan can swap at 32K; the retained path completes 32K,
  65K, and 100K with zero new swapout.
• GLM maps the exact norm/bias spans required by the compact-indexer transition.
  This fixes the former 32K failure without changing attention, KV geometry,
  RoPE, or context allocation.
• Acceptance benchmarks now require isolated 128, 2K, 8K, and 32K prompts,
  exact decode evidence, TPOT p50/p95, memory pressure, zero new swapout, and
  separate prose and security/coding lanes for routing/cache changes.
• CUDA and ROCm remain frozen with source and build targets absent. Distributed
  inference remains retired and its former flags fail closed.

M5 Pro 64 GB release evidence

Model and 32K workload	Prefill	Decode	Tail evidence
Qwen, resident prose final scalar source	65.76 t/s	about 3.18 t/s	315.193/317.518 ms p50/p95
DeepSeek, SSD prose final stack	164.43 t/s	7.27 t/s	105.410/149.943 ms p50/p95
GLM, SSD prose	44.73 t/s	1.87 t/s overall	472.624 ms p50, about 2.12 t/s steady

The DeepSeek original controls can swap, so this release does not publish a
32K speedup percentage. GLM also has no speedup percentage because the original
baseline fails before logits. Its same-prompt corrected output is identical to
the earlier gold and the final 32K run has zero swap. A separate security/coding
GLM prompt routes through more experts and measures 1.33 t/s; the difference is
workload locality, not a same-condition regression.

Full artifact identities, prompt hashes, invalid cohorts, 65K/100K DeepSeek
results, and functional API/agent/disk-KV checks are in
2026-07-20-long-context-metal-stack.md.

Upgrade boundary

There is no backward-compatibility promise for old model files. Remove old
startup flags and select a qualified v2 file explicitly. Qwen and DeepSeek may
resolve to different residency modes on different supported Macs; GLM resident
mode is rejected even on a larger host until it has its own qualification lane.