kimi-vl

kimi-vl (2026-07-19) — RETIRED 2026-07-27, NO LONGER SERVABLE

Removed from the box 2026-07-27. llama-swap entry deleted and the 18GB Q8_0

GGUF wiped from the HF cache (93GB→110GB free). Rationale: its only claim was

vision, and it was the *weakest* VLM here — 17/21 vision against PERFECT 21/21

from both [[bonsai-prism-fork]] and [[fable-711-27b]], while being unusable for

anything else (10/23 coding, zero tool-calling). Fully redundant, so it was pure

disk cost. Everything below is history; re-add from the git history of

llama-swap.yaml and re-pull the GGUF if it is ever wanted again.

**Cleanup checklist this retirement exercised — a model id lives in SEVEN places,

five of which no generator touches:** llama-swap.yaml (the entry) and

models.json + both statusline-command.sh copies are handled by

.claude/sync-models.sh, but you must hand-edit (1) that script's own

VISION_IDS set, (2) bin/gen_kilo_config.py — its model list, NO_TOOLS

set and DISPLAY dict — then re-run it, (3) ~/.config/opencode/opencode.json,

(4) ~/.kimi/config.toml, and (5) the local-model-router skill's table.

Leaving 3-5 stale makes those clients advertise a model that now 404s.

See [[sync-models-hook]], [[kilo-code-setup]], [[opencode-setup]],

[[kimi-cli-setup]].

moonshotai/Kimi-VL-A3B-Thinking-2506 via the OFFICIAL ggml-org GGUF (Q8_0

17GB + mmproj — a REAL vision tower, unlike [[agentworld-vision-stub]]).

16B/~3B-active MoE, DeepSeek-V3-lite/MLA text tower + MoonViT. Served as

kimi-vl on port 9111, c=131072 native, temp 0.8 (card), q8/q8 KV

(~10.1/11.0GB per GPU, ~5.5GB free each), ~80-97 t/s.

Two hard limitations found:

1. ◁think▷ markers stay in content — llama.cpp has no parser for

Kimi's ◁think▷…◁/think▷, so reasoning is never split into

reasoning_content. bench/grill_round3/4 clean_content was extended

(2026-07-19) to strip them; any other client must do the same.

2. No tool-call format at all — the 907-char template has no tools

section. Tools requests don't 400 (no jinja guard) but the model answers

in prose; structured tool_calls never appear. r4 agentic ran 0 tool steps.

Grill (temp 0.8): CODING 10/23 (r1 3/8, r2 3/5 — burns to the 16k cap

mid-think on hard tasks; r3 3/5, r4 1/5 per limitation 2). VISION 17/21

(easy 5/6, med 12/13, hard 0/2).

How to apply: niche only — small/fast vision with visible long-form

reasoning. vision-coder (21/21 vision, 22/23 coding) remains the vision

flagship; never route agentic/tool work to kimi-vl. -hf auto-pairs the

repo's Q8_0 mmproj. See [[vision-coder-heavy-grill]].