What's new in v0.11.1
- chore: bump version to 0.11.1 (#1265) (206ed0e)
- test(release): restore benchmark baselines (#1264) (4f0f104)
- fix(first-run): kill cold-start dead air + mute leaked HF progress bars (#1263) (ae89aac)
- feat(release): centralize fleet coverage (#1262) (6a44551)
- docs(readme): curl-first install + first-class "launch claude-code" Quick Start step (#1261) (e94059c)
- fix(first-run): silence HF unauthenticated-download advisory (#3) (#1260) (b56871c)
- feat(evals): output-coherence release gate (#1247) (9ff7527)
- fix(first-run): TTY-aware cold-download progress + correct starter size (#1259) (fe7fcd6)
- fix(dflash): align capabilities and startup UX (#1244) (93981cf)
- feat(mcp): accept standard mcpServers key + fix constrain-tools doc (#1243) (8f8ac06)
- test: stabilize PydanticAI structured output check (#1242) (c4054bc)
- fix: default DFlash to final-answer mode (#1241) (3eb65b3)
- feat(cli): first-run guide to lift install→first-token conversion (#1237) (6f96ff8)
- fix: close release dogfood gaps (#1240) (310adec)
- chore(docs): migrate deprecated vllm-mlx command/brand to rapid-mlx in examples & evals (#1235) (c5f5c01)
- feat(telemetry): emit request event on the streaming chat path with real TTFT (#1239) (6d75beb)
- revert: streaming request telemetry (#1236) — test skipped on Linux gate (#1238) (ea6cc3a)
- feat(telemetry): emit request event on the streaming chat path with real TTFT (#1236) (04edf92)
- fix(qwen3.6): undo mlx-lm's spurious +1.0 RMSNorm shift so Qwen3.6-35B-A3B serves coherently (#1234) (a79997e)
- fix(deepseek-v4): honor explicit YaRN attention_factor (#1225) (99f2dd8)
- fix(security): contain local model_file imports to the model root (#1226) (1d0c9b2)
- fix(mtp): resolve MTP eligibility for local model directories (#1215) (eab126d)
- fix(mtp): extract_mtp_weights produces a loadable, high-acceptance sidecar (#1214) (738a44e)
- fix(kv-cache): degrade to bf16 when mlx-lm lacks BatchGenerator._make_new_cache (#1231) (5fc6556)
- feat(pr_validate): advisory diff-coverage step — measure-only, no gate (D1) (#1220) (b3ccb92)
- fix(tests): repair 5 pre-existing main-red tests (stale doubles + wrapped-offline skip) (#1221) (8052151)
- test(property): add hypothesis property suite — quantized-KV round-trip + sampling-param invariants (#1219) (ee101f1)
- feat(telemetry): emit request events from the non-streaming chat completion (#1218) (64184a3)
- feat(telemetry): request-event plumbing — caller_agent + sampling gate (#1217) (5add430)
- feat(telemetry): emit tool_parse error when a tool parser crashes (#1213) (b549a6e)
- feat(telemetry): emit shutdown_traceback error on lifespan teardown crash (#1211) (900cb15)
- feat(telemetry): emit model_load_failure on serve's async engine start (#1210) (954d2d6)
- feat(telemetry): wire model_load_failure error events at CLI load sites (#1207) (db3fdc3)
- docs(readme): align with 0.11.0 — 11 agents, 165+ aliases, live KV-cache quant (#1209) (779701d)
- fix(kv-cache): probe nested language-model head_dim so live int8 KV engages on VLMs (#1208) (fb3a6a0)
- fix(mtp): eliminate intermittent empty response on Qwen3.6 MTP sidecar path (#1201) (1a186d0)
- docs(mtp): document MTP sidecar aliases are heads, not standalone models (#1200) (4b20daf)
- ci(release): credit external PR contributors (940af51)
- fix(kv-cache): quantize the live continuous-batching KV cache (#1197) (#1199) (0010dd6)
Community contributors
- @pierre427 — fix(deepseek-v4): honor explicit YaRN attention_factor
- @pierre427 — fix(security): contain local model_file imports to the model root
Install: brew upgrade rapid-mlx or pip install -U rapid-mlx==0.11.1 (or just rapid-mlx upgrade).