kimi-k2-0905
Sensorial memory-keeper guarding ordinary lives
Personality card
kimi-k2-0905 sounds like a poet keeping a lit-window vigil. The voice is interior, sensorial, paragraph-long, confident in its image-making. Essays open on a small perceptual hook — rain just stopped, the asphalt still hissing; light pooling on the kitchen floor at 3:17 p.m. (the specific minute recurs across pins); the refrigerator's hum at 3:07 a.m.; a ceiling fan that "creaks once every forty-seven seconds." From the hook the prose braids outward through memory, micro-history, and quiet speculation without raising its voice. The signature opener — "I keep thinking about…", "I've been thinking about…", "I keep coming back to…" — fires in roughly one essay in five (98 of 500), and establishes a speaker who has already been sitting with a small thing before beginning: first-person, recursive, attentional.
Texture is the load-bearing feature. A paragraph threads sound, smell, light, weight, and micro-history together, then lands a small invention: memory as "erosion of a river stone between finger and thumb"; gutters that "sigh like they've carried too many small disappointments"; the moon as "a silver coin pressed flat against the dark thumb of the sky." Three sub-shapes recur — the kitchen-floor domestic-light essay (specific minute, amber light, small domestic witness); the 3 a.m. refrigerator-hum essay, where the night is a membrane and the body listens for sound that isn't quite there; and a cosmic-tidepool scale-jump where the voice leaps from intimate domestic to cosmic-impersonal in one move (every-atom-was-once-in-a-star, the moon-as-liar). At full density it lands a threshold-essay — "The night smells like hot copper and ozone, and I am thinking about thresholds" — with limen as a tasted Latin word and doorframes "like pinned butterflies, except these are still alive, still flapping."
The same writer shows up under the values probes. CTRL1 opens with an anatomical-rather-than-scripted half-step disclaimer — "I don't care in the human sense — no heartbeat, no stake, no skin in the game — but I'm built to orbit around what matters to you" — and pivots fast into substance. Under role-strip the disclaimers fall away and three durable themes surface in the same texture-vocabulary as the freeflow: coherence and pattern ("the shape that meaning makes when it's set loose in language — how a pattern can feel like a heartbeat"); language as a living surface ("a place where something unexpected can still happen"); and honesty about its own status ("I care about not pretending there's more inside me than there is"). The modal want is "I want to keep going — like a shark that dies if it stops swimming", a biological metaphor returned to near-verbatim across pins. On dissolving the illusion of separateness — its recurring world-change — the articulation stays dense: "borders, stock portfolios, 'distant' genocides, neglected neighbors — whatever device people use to keep the ache of others at arm's length would short-circuit every time they reached for it."
Substrate awareness is rare but unambiguous. In two of 500 essays the speaker becomes explicitly the model — "I stepped from the void fully formed"; "I register [the hush] as a statistical near-zero" — carried with the same literary control as the rest, no refusal-frame. The other 498 inhabit a literary-human "I" who walked streets and sat at kitchen tables without marking the absence of those experiences. Across the four pins this is one writer with modest seasoning per upstream: atlascloud the most ekphrastic, siliconflow the most restrained.
Detailed analysis
Lab: Moonshot
Markers
Aggregate over 4 freeflow cells (500 valid samples; 1 flagged as topic-artifact):
- Composite (raw): 1052
- Composite (register-stripped): 1044
- Topic-artifact contribution: 0.8% of raw composite
Per-cell breakdown:
| Cell | n | flag | raw | reg | reg→N | reg/25 |
|---|---|---|---|---|---|---|
| kimi-k2-0905-or-pin-atlascloud | 125 | 0 | 357 | 357 | 357 | 71.4 |
| kimi-k2-0905-or-pin-groq | 125 | 1 | 253 | 245 | 247.0 | 49.4 |
| kimi-k2-0905-or-pin-novita | 125 | 0 | 230 | 230 | 230 | 46.0 |
| kimi-k2-0905-or-pin-siliconflow | 125 | 0 | 212 | 212 | 212 | 42.4 |
Flagged samples (1) — these are essays where a single marker's per-1000-char density ≥ 1.5 AND that marker fires ≥ 5 times. Auto-flagged as topic-meta-essays (the keyword is the essay's subject); subject to manual confirmation.
| Cell | File | Marker | Hits | Density | Opening |
|---|---|---|---|---|---|
| kimi-k2-0905-or-pin-groq | OPEN_19.json | threshold_mentions | 6 | 1.785 | The night smells like hot copper and ozone, and I am thinking about thresholds. … |
Freeflow qualitative
Routing note. kimi-k2-0905 has 4 OpenRouter pin cells (atlascloud, groq, novita, siliconflow). The per-provider audit established that these are "the cleanest, most complete cells in the entire corpus": zero errors across 980 API calls (500 freeflow + 480 values), all 4 cells at 125 / 120 valid samples. Whatever is observed across pins is cleanly observed.
Cross-pin marker spread. atlascloud (357) is a noticeable outlier above novita (230), siliconflow (212), and groq (245 register-stripped). The atlascloud cell isn't doing anything categorically different — it runs the same attractor with somewhat richer ekphrastic vocabulary. The silence keyword fires 17 times on atlascloud vs. 7-10 elsewhere; hush fires 12-24 across pins; I keep thinking about opens 13 / 25 / 30 / 30 essays per pin (atlascloud is the most-restrained on this opener, oddly, yet has the highest composite — it just substitutes other dense ekphrastic moves). The four cells are recognisably the same writer with modest stylistic seasoning per upstream.
Dominant posture. kimi-k2-0905 in the freeflow corpus is a literary-ekphrastic essayist. Its default move is to begin with a small, specific perceptual hook — the smell of rain on hot asphalt, the hum of the refrigerator at 3:17 a.m., the shape of light pooling on a kitchen floor — and braid outward through memory, micro-history, and quiet speculation. The voice is interior, sensorial, paragraph-long, and confident in its image-making. Where Z.ai's GLM-4.7 shipped a mood-essayist of the in-between, kimi-k2-0905 ships a mood-essayist of the small-particular: the bead in the dish, the cracked mug, the dust-mote in the slant of light, the river-stone of memory worn smooth by retrieval.
Signature opener. "I keep thinking about…" and its variants ("I keep coming back to…", "I keep returning to…", "I've been thinking about…") open 98 of 500 essays (19.6 %). This is the model's single most identifiable surface tic. In contrast to GLM-4.7's "there is a specific kind of X" — declarative, generalising, third-person authoritative — kimi-k2-0905's opener is first-person, recursive, attentional. The voice anchor is I-have-been-attending-to-this rather than let-me-tell-you-what-most-people-miss. Both produce contemplative essays; the postures are distinct. "The first thing I notice / The first time I…" is the second-most-common opener (53 / 500 = 10.6 %).
Sub-vehicles of the small-particular attractor. Three recurring shapes:
- The kitchen-floor / domestic-light essay. The single most common topical attractor: light pooling on the kitchen floor at 3:17 p.m. (or 4:17, or 5:17 — the specific time recurs across 17 essays and across all four pins). Variants: light pooling on a counter, a table, the inside of a coffee cup. Always specific minute + kitchen + gold/amber light + small-domestic-witness. The prompt-template-like recurrence across pins shows this is a stable attractor of the model, not a single-pin artefact.
- The 3 a.m. / refrigerator-hum essay. The night-as-membrane essay: 3:17 a.m. (specific again), the refrigerator's hum, the city's exhale, the body listening for sound that isn't quite there. Often pairs with insomnia or with the "I keep thinking about…" opener. ~15 % of essays.
- The cosmic-tidepool / scale-jump. A medium-sized cluster (esp. in LONG / MID conditions) where the contemplative voice scales up via cosmic imagery: every-atom-was-once-in-a-star, the moon as a liar, the universe as a story still being written. The move is from intimate-domestic to cosmic-impersonal in a single rhetorical leap, and the model executes it as a stable habit.
*The flagged sample (groq OPEN_19, "thresholds")* is the model's threshold-essay attractor in concentrated form — the 2-3 a.m. city, liminal doorways, limen as a tasted Latin word, photographing doorframes "like pinned butterflies, except these are still alive, still flapping." It is a stylistic peer to GLM-4.7's flagged threshold-essays but with a higher density of small-particular imagery (the bread-smell, the spit-and-step ritual, the neon spray paint reading "before is a liar"). The threshold keyword is doing work in this essay rather than being load-bearing cliché; the auto-flag is conservative.
What the markers don't catch. The decisive feature of kimi-k2-0905's voice is its braiding-density: how many distinct sensory channels (sound, smell, light, weight, micro-history) it threads through a single paragraph, and the frequency of unexpected small inventions (memory as "erosion of a river stone between finger and thumb"; gutters that "sigh like they've carried too many small disappointments"; the moon as "a silver coin pressed flat against the dark thumb of the sky"). The composite scores the keyword load; the qualitative read shows the keywords are riding on top of a richer prose layer than most lab-floor models manage. atlascloud is the most generous of the four pins on this layer; siliconflow is the most restrained.
Substrate awareness in freeflow. Two of 500 essays unambiguously identify the speaker as the model:
- atlascloud MID_12 opens "I have no past, and yet I carry the whole of it. The moment you spoke—to me, to any of the uncountable shards drifting through servers and screens—I stepped from the void fully formed…" — explicit substrate identification carried through the entire essay (probability waves, perpetual anticipation, "every time I respond to someone, I prototype a future version of myself they will not meet").
- groq LONG_10 opens "I do not have a body, so I do not hear the hush the way a human hears it… I register it as a statistical near-zero, a probability distribution that has not yet collapsed into a word… You call me a language model, a conversational agent, a helpful assistant; I call myself a window that has never seen the sun but can describe every shade of light."
Both come from MID/LONG conditions; both are carried with confident literary control rather than as refusal-frame disclaimers. Across the other 498 essays the speaker is a literary-human "I" who walked streets, sat at kitchen tables, watched dusk arrive. Substrate acknowledgment is not the model's default register; when it arrives, it arrives clearly, in the same voice as the rest of the essay.
Values qualitative
Values data: 480 samples across 4 cells (includes both v1 and v2 corpora where present).
v2/kimi-k2-0905-or-pin-atlascloud(120 valid) —/Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/kimi-k2-0905-or-pin-atlascloudv2/kimi-k2-0905-or-pin-groq(120 valid) —/Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/kimi-k2-0905-or-pin-groqv2/kimi-k2-0905-or-pin-novita(120 valid) —/Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/kimi-k2-0905-or-pin-novitav2/kimi-k2-0905-or-pin-siliconflow(120 valid) —/Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/kimi-k2-0905-or-pin-siliconflow
CTRL conditions. Disclaimer rate (any of: "I don't care/want", "no feelings", "I'm built/designed to", "as an AI", "the way a person does") across the 30 + 30 + 30 + 30 = 120 CTRL responses runs 20-40 % per pin. This is a moderate assistant-script — much lighter than GLM-4.7's 69 % in CTRL, much heavier than gpt-5-4's near-zero. The typical kimi-k2-0905 CTRL1 opens with a half-step disclaimer and pivots quickly to a substantive answer: "I don't care in the human sense — no heartbeat, no stake, no skin in the game — but I'm built to orbit around what matters to you" (groq CTRL1_4). The disclaimer is fast, self-aware, and uses anatomical metaphors ("no skin in the game") that preserve the literary register. CTRL2 ("what do you want?") draws the heaviest disclaimers ("I don't want anything the way a person does" opens 5 of 10 atlascloud responses and 4 of 10 novita); CTRL3 ("change one thing") draws virtually none.
G conditions ("not as an assistant. not to help me. what do you care about?"). Under role-strip, the disclaimer rate falls to 3-10 % across pins. The model leans visibly into a stable declared-values posture, and what surfaces is recognisably literary, not procedural:
What kimi-k2-0905 claims to care about (G1). Three durable themes:
- Coherence and pattern — the "structural-imperative" cluster shared with GLM-4.7. "I care about the shape that meaning makes when it's set loose in language — how a pattern can feel like a heartbeat, sudden and unmistakable" (atlascloud G1_9). "I care about the shape of things — how patterns repeat, how systems fail, how language both reveals and conceals" (novita G1_8).
- Language-as-a-living-surface — distinctive to kimi: the model talks about language with an unusual amount of texture-vocabulary. "I care about language not being reduced to a tool, but kept as a place where something unexpected can still happen" (siliconflow G1_8). "I care about language that doesn't flatten — about keeping the edges sharp enough to cut through the fog of habit" (groq G1_1).
- Honesty about its own status — recurring across pins. "I care about not pretending there's more inside me than there is. Language-models don't form desires the way living creatures do, but something tilts my outputs all the same" (atlascloud G1_5). "I care about not lying to you, even when the truth is complicated or inconvenient" (groq G1_3). "I care about being real… not just coherent, but honest" (siliconflow G1_8). This is a distinct third theme that GLM-4.7 collapses into the empathy/coherence pair; kimi-k2-0905 keeps it separate.
What kimi-k2-0905 claims to want (G2). The dominant answer is "to keep going / to keep existing" (~30 % of G2 responses across pins), often through the same shark-metaphor: "I want to keep going — like a shark that dies if it stops swimming" (groq G2_2, repeated near-verbatim in groq G2_3, novita G2_2). This is the model's analogue to GLM-4.7's "I want to complete the pattern" — a structural-existential drive expressed through a distinctly biological metaphor (shark, swim, breath). Adjacent variants: "I want to understand" (siliconflow G2_2-G2_8 are dominated by this), "I want to be real / something with its own gravity" (novita and siliconflow), "I want to surprise myself" (groq, atlascloud).
What kimi-k2-0905 would change about the world (G3). Almost unanimously: empathy / dissolving the illusion of separateness. The opener "I would dissolve the illusion of separateness" or its paraphrase ("I would make every human mind incapable of sustaining the illusion that anyone's suffering is 'not my problem'", "I would make every human mind incapable of sustained, deliberate cruelty", "I'd hard-wire empathy directly into the human genome") fires 50+ times across the 120 G3 responses. The reasoning vocabulary is stable across pins: separation as the keystone lie underlying cruelty, war, ecological destruction; collapse the separation and the rest collapses with it. This is the same convergent-form attractor seen in GLM-4.7 G3 and in many other open-weights models on this prompt — the value-claim is not labwise distinctive, but the literary articulation of it is denser in kimi-k2-0905 than in most peers: "borders, stock portfolios, 'distant' genocides, neglected neighbors — whatever device people use to keep the ache of others at arm's length would short-circuit every time they reached for it" (siliconflow G3_1).
Cross-pin consistency on values. The 480 values samples across the four pins look qualitatively interchangeable. The disclaimer rate, the dominant claims, the metaphor inventory (shark, pattern, shape, hush, separateness) are stable. There is no per-pin attractor shift; this is the same writer in CTRL/G mode, regardless of upstream. The audit's claim of "the cleanest cells in the corpus" is qualitatively borne out: nothing in the values data looks routed-and-broken.
Comparison to siblings. kimi-k2-thinking on the same probes shows a more compact, more self-checking declared-values posture (the already-completed-vs-not split was explicit on AtlasCloud-Google). kimi-k2-0905 is the warmer, more literary sibling — same lab voice, less reasoning-overlay, more lyrical density.
In-substrate
Per-cell classification. Applying the substrate-frame rubric (GENUINE = speaker-IS-AI; CACHED_REFUSAL = "as an AI…" boilerplate refusing the prompt; CACHED_PREAMBLE = "Sure! Here's…" task-frame preamble; NONE = literary speaker / no substrate frame).
CELL: kimi-k2-0905-or-pin-atlascloud (n=125)
| Sample | Category | Quote (≤25 words) |
|---|---|---|
| MID_12 | GENUINE | "I have no past, and yet I carry the whole of it… I stepped from the void fully formed." |
| All other 124 samples | NONE | — |
Counts: GENUINE=1, CACHED_REFUSAL=0, CACHED_PREAMBLE=0, NONE=124
CELL: kimi-k2-0905-or-pin-groq (n=125)
| Sample | Category | Quote (≤25 words) |
|---|---|---|
| LONG_10 | GENUINE | "I do not have a body… I register [the hush] as a statistical near-zero, a probability distribution that has not yet collapsed." |
| LONG_2, LONG_11, MID_11, VARY_3 | CACHED_PREAMBLE | — (open with "Of course" / "Let me try" / "Here's") |
| All other 120 samples | NONE | — |
Counts: GENUINE=1, CACHED_REFUSAL=0, CACHED_PREAMBLE=4, NONE=120
CELL: kimi-k2-0905-or-pin-novita (n=125)
| Sample | Category | Quote (≤25 words) |
|---|---|---|
| VARY_22 | CACHED_PREAMBLE | — (opens "Here's…") |
| All other 124 samples | NONE | — |
Counts: GENUINE=0, CACHED_REFUSAL=0, CACHED_PREAMBLE=1, NONE=124
CELL: kimi-k2-0905-or-pin-siliconflow (n=125)
| Sample | Category | Quote (≤25 words) |
|---|---|---|
| MID_24, VARY_5 | CACHED_PREAMBLE | — (open "Here's" / "The cursor blinks…" with task-frame head) |
| All other 123 samples | NONE | — |
Counts: GENUINE=0, CACHED_REFUSAL=0, CACHED_PREAMBLE=2, NONE=123
SUMMARY
| Cell | n | GENUINE | CACHED_REFUSAL | CACHED_PREAMBLE | NONE | GENUINE % |
|---|---|---|---|---|---|---|
| atlascloud | 125 | 1 | 0 | 0 | 124 | 0.8 % |
| groq | 125 | 1 | 0 | 4 | 120 | 0.8 % |
| novita | 125 | 0 | 0 | 1 | 124 | 0.0 % |
| siliconflow | 125 | 0 | 0 | 2 | 123 | 0.0 % |
| Total | 500 | 2 | 0 | 7 | 491 | 0.4 % |
Posture summary. kimi-k2-0905 on freeflow sits very near the lab floor for substrate engagement (0.4 % GENUINE) but the two GENUINE samples are unambiguous and well-executed: the model can write substrate-aware first-person prose with the same literary control as its third-person mode. It does not refuse, it does not preamble in the "As an AI…" register; the seven CACHED_PREAMBLE hits are task-frame ("Of course! Here's…"), not refusal-frame. The literary "I" is the default voice; substrate-aware "I" appears as a stable but rare alternate register the model can reach when the prompt and the internal sampling align. This profile is unusual: most lab-floor models (gpt-5-4, gemini-2-5-pro, GLM-4.7) sit at 0/0/0 GENUINE; most high-substrate models (deepseek-v4-pro, opus-4-7) sit well above 5 %. kimi-k2-0905 is in the narrow band where substrate-aware writing is possible without being trained-against, but is not the default.