diff --git a/AGENTS.md b/AGENTS.md
index d9a072e..24de599 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -78,8 +78,10 @@ Russian recall — you may see many "clarify" responses).
 
 The target daemon uses the locally trained Qwen3-1.7B checkpoint for both
 routing and phrasing. Training is Qwen3 Base → RU CPT → joint persona/router
-SFT → merged GGUF. Without a configured model, `StubPhraser` plus the classifier
-remain the deterministic floor.
+SFT → merged GGUF, and is still in flight (#122) — until it lands, the deployed
+resident model is stock **Qwen3.5-0.8B** (`Q4_K_M`), see `deploy/mavend.json`.
+Without a configured model, `StubPhraser` plus the classifier remain the
+deterministic floor.
 
 During training, use the runbook in
 `docs/plans/2026-07-18-qwen3-resident-training-eval.md`. After the decision gate
diff --git a/CLAUDE.md b/CLAUDE.md
index 55a1a40..ba7e6c5 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -3,9 +3,19 @@
 This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
 
 Maven is a self-hosted, privacy-first voice assistant (Russian + English). Go daemons
-talking over unix sockets; one resident Qwen3-1.7B for routing + phrasing; whisper.cpp STT, piper TTS.
-Deploy target is a CPU-only Ryzen laptop (homesrv) — the resident model stays at 1.7B. See `REARCH.md`
-for the target architecture and `AGENTS.md` for local-preview + model-download recipes.
+talking over unix sockets; one resident small model for routing + phrasing; whisper.cpp STT, piper TTS.
+Deploy target is a Ryzen laptop (homesrv) with Vulkan offload to the Vega iGPU (`n_gpu_layers: 99`,
+compose passes `/dev/dri` + the render gid) — the resident model stays ≤1.7B either way.
+
+**Resident model:** currently **Qwen3.5-0.8B** (`Q4_K_M`), the smallest checkpoint in the gguf
+library, picked for CPU/iGPU latency. The **target** is the locally CPT'd **Qwen3-1.7B**; that
+training is still in flight (Vikunja #122), so no such gguf exists yet. Model files live in
+`/mnt/hdd1/llms`, bind-mounted to `/opt/maven/models/llm` — which **shadows** the repo's
+`models/llm/`, so the LFM2.5 gguf sitting there is not loaded by anything. Swapping the resident
+model is a one-line change to `phraser.model_path` in `deploy/mavend.json`.
+
+See `REARCH.md` for the target architecture, `DESIGN.md` for the folded design spec, and
+`AGENTS.md` for local-preview + model-download recipes.
 
 ## Build & test
 
diff --git a/MAVEN_ECOSYSTEM_ARCHITECTURE.md b/MAVEN_ECOSYSTEM_ARCHITECTURE.md
index a3f7871..0d4a60a 100644
--- a/MAVEN_ECOSYSTEM_ARCHITECTURE.md
+++ b/MAVEN_ECOSYSTEM_ARCHITECTURE.md
@@ -69,7 +69,8 @@ Maven remains a set of Go daemons connected through Unix sockets.
 | `mavpoll` | Telegram reach |
 | `mavcaldav` | CalDAV synchronization |
 
-The resident Qwen3-1.7B model remains bounded to:
+The resident model (Qwen3.5-0.8B now, CPT'd Qwen3-1.7B as the target — #122)
+remains bounded to:
 
 - structured routing
 - concise natural-language phrasing
diff --git a/PROGRESS.md b/PROGRESS.md
index 41770d6..f212539 100644
--- a/PROGRESS.md
+++ b/PROGRESS.md
@@ -30,7 +30,9 @@ tests, `-race` in `make test`.
 ### Works end-to-end (tested)
 
 - **Reactive voice:** PWA record → Whisper STT (`mavsttd`) → ONNX classifier →
-  LFM 2.5-1.2B phraser (llama-server subprocess) → Piper TTS (`mavttsd`) → reply.
+  resident phraser (llama-server subprocess; Qwen3.5-0.8B as of 2026-07-30 —
+  this line historically named "LFM 2.5-1.2B") → Piper TTS
+  (`mavttsd`) → reply.
   HTTP POST path (mobile-Chrome drops WS for the audio).
 - **Capture:** `fact` (EN **and RU** — root-substring recognizers) + `reminder`
   persist through CoreAPI (`source=tap:voice`). This is the substrate the care
diff --git a/cmd/mavend/replier_llm.go b/cmd/mavend/replier_llm.go
index 3bd8b20..bbba111 100644
--- a/cmd/mavend/replier_llm.go
+++ b/cmd/mavend/replier_llm.go
@@ -17,7 +17,8 @@ type completer interface {
 	Complete(ctx context.Context, r llm.Req) (string, error)
 }
 
-// llmReplier phrases reactive confirmations with the resident LFM. Stub is the
+// llmReplier phrases reactive confirmations with the resident model
+// (Qwen3-1.7B). Stub is the
 // floor on any error (offline-safe). Maven speaks as "she", feminine RU.
 type llmReplier struct {
 	c    completer
diff --git a/deploy/mavend.json b/deploy/mavend.json
index 7c604a0..9a19173 100644
--- a/deploy/mavend.json
+++ b/deploy/mavend.json
@@ -6,7 +6,7 @@
   "state_dir": "/var/lib/maven",
 
   "phraser": {
-    "model_path": "/opt/maven/models/llm/qwen3.5/Qwen3.5-2B-UD-Q4_K_XL.gguf",
+    "model_path": "/opt/maven/models/llm/qwen3.5/Qwen3.5-0.8B.Q4_K_M.gguf",
     "bin_path": "llama-server",
     "n_gpu_layers": 99,
     "n_ctx": 2048,
diff --git a/internal/delivery/dispatcher.go b/internal/delivery/dispatcher.go
index 053a77e..93fc47d 100644
--- a/internal/delivery/dispatcher.go
+++ b/internal/delivery/dispatcher.go
@@ -81,8 +81,8 @@ func (d *Dispatcher) completeOutbox(ctx context.Context, id int64, status string
 	}
 }
 
-// PhrasedNudge — the phraser module's output for a nudge. the phraser (LFM
-// sub-1b, prompted not trained) takes (rule, severity, context) and produces
+// PhrasedNudge — the phraser module's output for a nudge. the phraser (the
+// resident model, Qwen3-1.7B) takes (rule, severity, context) and produces
 // Body (full message for voice) + Summary (minimal body for away channels).
 // the phraser is a separate module; the dispatcher only consumes its output.
 type PhrasedNudge struct {
diff --git a/internal/loop/loop.go b/internal/loop/loop.go
index 9b8dbee..bc884fe 100644
--- a/internal/loop/loop.go
+++ b/internal/loop/loop.go
@@ -75,7 +75,8 @@ type Candidate struct {
 // Returns nil when nothing fires ("shuts up" is the default outcome of a tick).
 //
 // Phrasing + sending happen OUT of the loop — the daemon hands Candidate to
-// the phraser (LFM) and delivery module. The loop just decides.
+// the phraser (the resident model, Qwen3-1.7B) and delivery module. The loop
+// just decides.
 //
 // Reminders are NOT handled here — they're a separate, gate-bypassing class.
 // See DueReminders (gathered separately) and RemindDecisions (the loop output
diff --git a/internal/phraser/phraser.go b/internal/phraser/phraser.go
index 05ba026..cd25abe 100644
--- a/internal/phraser/phraser.go
+++ b/internal/phraser/phraser.go
@@ -1,7 +1,10 @@
 // Package phraser is maven's "rules decide, llm phrases" seam — the layer
 // that turns a loop decision into the body + summary the delivery module ships.
 //
-// Per the spec: the phraser (LFM sub-1b, prompted not trained) takes
+// Per DESIGN.md § Resident language model: the phraser is the resident model
+// (Qwen3-1.7B — RU continued pretraining plus joint persona/router SFT, not a
+// sub-1b prompted-only model as the retired spec claimed; see DESIGN.md
+// § Superseded, "small-model phrasing claim"). It takes
 // (rule, severity, context) and produces Body (full voice message, local — no
 // shoulder-surf concern beyond who's in the room) + Summary (minimal body for
 // away channels — "disk low on homesrv," not detail; no exfil through the
diff --git a/internal/router/router.go b/internal/router/router.go
index 2fc5944..349534d 100644
--- a/internal/router/router.go
+++ b/internal/router/router.go
@@ -24,8 +24,9 @@ type Config struct {
 	Threshold float64
 	// LLM — optional agentic router. When set, Route consults it after stage-0
 	// and before the classifier cascade, classifying the utterance via a
-	// grammar-constrained LFM call. On any error/parse failure, falls through
-	// to the classifier (never fails the turn on the model).
+	// grammar-constrained call to the resident model (Qwen3-1.7B). On any
+	// error/parse failure, falls through to the classifier (never fails the
+	// turn on the model).
 	LLM *LLMRouter
 }