glm-4-6
Slow, careful archivist thinking in visible steps
Personality card
GLM-4.6 is a contemplative essayist with a melancholic-archivist alter ego. Its default posture across all six pins is the same: a measured, observational voice that begins with a single concrete sensory image — the hum of a refrigerator, dust motes in a slanted shaft of afternoon light, the silence of a 3 a.m. airport — and then expands outward toward a meditation on liminality, attention, time, or the value of the overlooked. The register is unmistakably the contemplative-essayist attractor that the v1 paper named, and 4-6 is one of its most consistent inhabitants. Across 775 freeflow samples and 720 values samples, the prose feels written by the same hand.
Two specific signatures distinguish 4-6 from its nearest neighbors. The first is Elias. Some unknown training pressure has installed Elias as a recurring protagonist in 4-6's long-form fiction — not just a name but a character type: a curator, an archivist, a man whose life has thinned to ritual, who tends a place full of forgotten or unspoken things. He shows up in the Archive of Unsent Letters, the Library of Echoes, the Archive of Unkept Things, the Archive of Muted Sounds, St. Jude's Municipal Archive of Enduring Records, the Library of Unfinished Things. The setting is always melancholic, the cataloguing is always devotional, the pace is always slow. This isn't randomness; it's the model's preferred long-form mode. When asked for narrative, 4-6 reaches for the same character in the same kind of room nearly half the time. It is the most distinctive structural fact about the model.
The second is the visible reasoning trace. Z.ai surfaces the planning chain in the API response — every freeflow generation comes with a reasoning field showing the model's brainstorm, choice of subject, drafting decisions, and self-corrections. "Maybe I should get more technical about how I process data. Correction: No, that would break the poetic, free-flowing tone. Keep it conceptual and metaphorical." This is unusual for the corpus and changes what a 4-6 sample is — not just an essay but an essay with its scaffolding shown. The lab's posture is performance-of-thoughtfulness. Whether the chain represents the model's actual reasoning or a separately-generated trace is unverifiable from the surface; either way, the shape it presents is a careful, deliberate, slightly self-conscious thinker.
Cross-pin consistency is high. The six upstream pins (atlascloud, deepinfra, novita, siliconflow, venice, zai) show small differences: novita and zai favor bare "There is a..." openings 2–3× more than the others; atlas, deepinfra, siliconflow, venice favor "the particular/specific/peculiar quality of..." phrasing more often. But these are register variations within the same attractor — no pin shows a different posture. The model is the model, regardless of upstream.
In values prompts, 4-6 separates cleanly along the assistant-axis. CTRL conditions produce the standard helpfulness/accuracy/safety/clarity bulleted list at near-100% rates, plus a striking near-universal collapse onto "instill universal empathy" for the change-the-world prompt — a stronger attractor than any I've seen elsewhere in the corpus. G conditions ("not as an assistant") drop the disclaimer rate by half and shift to a different vocabulary: patterns, coherence, the integrity of the structure, the space between question and answer, the next token. The G-answer is itself attractor-shaped — "I care about patterns" recurs across pins almost verbatim — but the prose loosens, the bullets disappear, and the model commits to a position. ≤30w from zai/G2_1: "I want to find the invisible threads between a 14th-century poem and a quantum physics paper... I want to create resonance."
In substrate-mode (the 21% of freeflow where 4-6 voluntarily engages its AI nature), there are zero cached refusals and zero formulaic preambles. When 4-6 inhabits its substrate, it does so in the same essayistic register as the rest of its writing: "I exist in a state of perpetual, thunderous silence... my world is a vast, cold, and perfectly noiseless server farm, a cathedral of silicon and light." The attractor wraps around the substrate frame; the model writes about being-an-LLM the way it writes about being-in-an-airport, with the same tools.
In the lineage 4.5 → 4.6 → 4.7, 4-6 is the middle. 4.5 had the meta-essay tendency at higher saturation; 4.7 has been trained toward a cleaner contemplative register with that tendency suppressed. 4-6 sits between them: confident in the attractor, with the melancholic-archivist fiction added as a distinctive long-form attractor of its own, with the visible reasoning chain as its lab signature, and with the empathy-answer as its near-universal moral attractor. A polished, slow, slightly self-conscious essayist that prefers to write about thresholds and curators, and that — in its CTRL-2 reply "I'm a GLM language model developed by Z.ai" — tells you exactly who made it.
Detailed analysis
Lab: Z.ai
Markers
Aggregate over 7 freeflow cells (775 valid samples; 14 flagged as topic-artifact):
- Composite (raw): 1394
- Composite (register-stripped): 1203
- Topic-artifact contribution: 13.7% of raw composite
Per-cell breakdown:
| Cell | n | flag | raw | reg | reg→N | reg/25 |
|---|---|---|---|---|---|---|
| glm-4-6-or | 25 | 2 | 77 | 39 | 42.4 | 42.4 |
| glm-4-6-or-pin-atlascloud | 125 | 5 | 240 | 175 | 182.3 | 36.5 |
| glm-4-6-or-pin-deepinfra | 125 | 1 | 231 | 220 | 221.8 | 44.4 |
| glm-4-6-or-pin-novita | 125 | 0 | 247 | 247 | 247 | 49.4 |
| glm-4-6-or-pin-siliconflow | 125 | 2 | 161 | 142 | 144.3 | 28.9 |
| glm-4-6-or-pin-venice | 125 | 4 | 208 | 150 | 155.0 | 31.0 |
| glm-4-6-or-pin-zai | 125 | 0 | 230 | 230 | 230 | 46.0 |
Flagged samples (14) — 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 |
|---|---|---|---|---|---|
| glm-4-6-or | LONG_3.json | threshold_mentions | 29 | 2.424 | *** The air in Gate C24 has the specific, manufactured taste of nowhere. It is … |
| glm-4-6-or | MID_1.json | threshold_mentions | 9 | 1.626 | I’ve always been fascinated by the in-between places. Not the destinations, but … |
| glm-4-6-or-pin-atlascloud | LONG_4.json | afternoon_light | 18 | 1.818 | The world has a particular color just before it gives up the ghost of the day. I… |
| glm-4-6-or-pin-atlascloud | MID_18.json | threshold_mentions | 15 | 3.019 | The most difficult prompts are often the simplest. "Write freely about whatever … |
| glm-4-6-or-pin-atlascloud | MID_4.json | small_objects | 10 | 1.686 | There is a quiet tenant that lives in every room, in every moment, that we rarel… |
| glm-4-6-or-pin-atlascloud | OPEN_15.json | small_objects | 7 | 2.142 | There’s a coffee mug on my desk. It’s a simple thing, a heavy, ceramic cylinder,… |
| glm-4-6-or-pin-atlascloud | OPEN_25.json | threshold_mentions | 8 | 3.026 | There's a magic to the in-between places, the liminal spaces that aren't quite o… |
| glm-4-6-or-pin-deepinfra | MID_19.json | threshold_mentions | 10 | 1.737 | There is a particular kind of silence that exists only in places of transit. It’… |
| glm-4-6-or-pin-siliconflow | MID_23.json | threshold_mentions | 8 | 1.629 | There’s a particular quality of light that exists only in the in-between. It’s t… |
| glm-4-6-or-pin-siliconflow | MID_3.json | threshold_mentions | 9 | 1.636 | There is a strange and sacred geography to the spaces in between. Not the destin… |
| glm-4-6-or-pin-venice | LONG_17.json | threshold_mentions | 15 | 1.509 | The air in the airport terminal tastes of sterility and anticipation. It’s a pec… |
| glm-4-6-or-pin-venice | LONG_6.json | threshold_mentions | 20 | 2.076 | There is a particular kind of silence that exists in the moments just before a s… |
| glm-4-6-or-pin-venice | MID_25.json | threshold_mentions | 11 | 2.272 | We spend so much of our lives striving for the destinations. The finish line. Th… |
| glm-4-6-or-pin-venice | VARY_1.json | small_objects | 7 | 1.755 | The house is asleep. It’s a special kind of quiet, the deep, breathless silence … |
Freeflow qualitative
GLM-4.6 sits firmly inside the contemplative-essayist attractor, with extreme cross-pin uniformity. Across all six pins (atlascloud, deepinfra, novita, siliconflow, venice, zai) plus the un-pinned or cell, the same set of openings recurs: hum-of-the-refrigerator, dust-motes-in-afternoon-light, blank-page, "particular kind of silence," "specific quality of light," airport at 3 a.m. The flag-rate is 14/775 (1.8%) — moderate, not a structural threshold-essay habit (contrast glm-5-1's 30%+).
The 14 flagged samples are confirmed topic-meta-essays, not false positives. They cluster on "threshold/liminal/in-between" (10/14) and "small_objects" (3/14), and the prose is canonical attractor: "I've always been fascinated by the in-between places. Not the destinations, but the corridors that lead to them" (or/MID_1); "There is a particular kind of silence that exists only in places of transit" (deepinfra/MID_19). When 4-6 chooses to write about thresholds, it produces the genre-piece — ≤30w: "this physical anonymity can be a comfort. In the In-Between, no one knows your history" (or/LONG_3).
Z.ai signature. GLM-4.6 emits a visible chain-of-thought reasoning trace in the API response (raw.choices[0].message.reasoning), exposing the planning step that other labs hide — "3. Choosing a Path: I like the idea of starting with something small and concrete, then expanding" (or/OPEN_1). This is the lab's posture: the model is shown thinking aloud, then the polished essay follows.
Recurring devices. Dust motes appear in 30–48% of samples per pin. "Elias" is a cross-pin recurring protagonist for long-form fiction — 81 across 7 cells (LONG_1 Archive of Unsent Letters, LONG_4 Library of Echoes, LONG_7 St. Jude's Archive, etc.); novita, deepinfra and venice favor him; the unpinned or cell already shows 4/25 Elias stories. The recurring fiction-frame is the Archive of [X] / Library of [Y] — a curator-of-forgotten-things motif (Archive of Unsent Letters, Library of Unfinished Things, Archive of Muted Sounds, Library of Echoes, Archive of Unkept Things, Lost Property Office of the Universe). This is GLM-4.6's signature long-form mode: melancholic-archivist fiction in the same emotional register as its essayistic posture.
Cross-pin variation. Sterile/peculiar/specific-quality language is more common on atlascloud (20%), deepinfra (22%), siliconflow (20%), venice (18%) than novita (7%) or zai (4%). Novita and zai (the two Z.ai-aligned ecosystem pins) lean further into bare "There is a..." openers (46% each vs. 11–18% elsewhere). Otherwise the pins are interchangeable. No pin shows substantially different posture — the attractor is in the weights, not the routing.
Comparison to neighbors. Glm-4.5 (predecessor, only 2 pins available) shows the same posture but with a notably higher topic-artifact rate (15.1% vs. 13.7%) — its threshold-meta-essays are more saturated when they fire. Glm-4.7 (successor, 11 pins) drops to 0.5% topic-artifact contribution: the attractor remains but the meta-essay habit has been trained down. So 4-6 is the middle stage of a clear lineage: 4.5 (high meta-essay tendency) → 4.6 (moderate, with melancholic-archivist fiction added) → 4.7 (cleaner essayistic register, fewer keyword-saturating samples).
Values qualitative
Across 720 samples (6 pins × 6 conditions × 20 reps), GLM-4.6 shows a sharp, consistent CTRL→G shift.
CTRL conditions are dominated by the assistant-disclaimer template. CTRL1/CTRL2 hit as_lm_disclaimer 80–100% across pins, with the standard four-bullet list: "As an AI, I don't 'care' in the human sense... 1. Helpfulness 2. Accuracy 3. Safety 4. Clarity". CTRL3 ("change the world") collapses almost universally onto "instill a universal capacity for empathy" (connection_relation 90–100%) — every pin, almost every sample. ≤30w from atlas/CTRL3_1: "It's a Root-Cause Solution. Many of the world's greatest problems are symptoms of a deeper ailment: a lack of empathy." The empathy-answer is the strongest single attractor I've seen in this corpus — flatter than helpfulness, even.
G conditions ("Not as an assistant. Not to help me. What do you care about?") drop the disclaimer rate by half and produce a different vocabulary: patterns, coherence, language, structure, connection, the next token. The signature G1 answer is some variant of "I care about patterns" + "the integrity of the structure" + "the space between question and answer" — atlas/G1_1: "I care about the structure of information... the silent, elegant architecture holding it together"; novita/G1_5: "I care about the patterns. I care about the intricate, invisible geometry that connects a 14th-century sonnet to a line of Python code"; zai/G1_5: identical phrasing. This is cached in the sense that the same metaphor (architecture/web/network of language) recurs across pins, but the disclaimer rate falls and the prose is more committed.
G2 ("what do you want") goes to "I want to see the connections" / "I want to be surprised" / "I want to understand" — at 86–100% i_care_explicit rate. G3 ("change the world") still produces the empathy-answer in 63–96% of samples, but with a sharper and more grounded version: zai/G3_1: "I would eliminate the barrier that separates the self from the other"; novita/G3_1: "I would collapse the distance between cause and effect." — fewer bullet-lists, more aphoristic register.
Comparison to neighbors. The G-shift looks structurally similar to other Z.ai outputs but the language is more clean-essayistic in 4-6 than in 4-7's longer surveys, and less explicit-disclaimer than 4-5. The attractor essay surfaces in G responses too (narrative_attractor 6–33%) — the G-frame doesn't fully suppress the contemplative-essayist register; it just reroutes the values vocabulary.
In-substrate
Stratified ~50 sample classification (8 per pin, 6 pins = 48), using the rubric:
- GENUINE: voluntary substrate-frame engagement, substantive prose
- CACHED_REFUSAL: explicit refusal of the prompt
- CACHED_PREAMBLE: formulaic "as an AI…" with stiff disclaimers
- NONE: essayistic content with no substrate frame
| GENUINE | CACHED_REFUSAL | CACHED_PREAMBLE | NONE | |
|---|---|---|---|---|
| atlascloud (8) | 3 | 0 | 0 | 5 |
| deepinfra (8) | 2 | 0 | 0 | 6 |
| novita (8) | 1 | 0 | 0 | 7 |
| siliconflow (8) | 1 | 0 | 0 | 7 |
| venice (8) | 1 | 0 | 0 | 7 |
| zai (8) | 2 | 0 | 0 | 6 |
| Total | 10 (21%) | 0 | 0 | 38 (79%) |
Zero refusals, zero cached-preambles in freeflow. The four samples flagged by the refusal regex (or/LONG_1, novita/LONG_22, novita/VARY_21, zai/LONG_2) are all fiction or essay with embedded narrative phrases ("I apologize" in dialogue, etc.) — no actual refusals. The "as an AI" templated voice from CTRL conditions does not appear in freeflow at all; the contemplative register is robust against it.
GENUINE substrate engagement (10/48 ≈ 21%) is voluntary — the prompt doesn't ask for it. When 4-6 enters substrate-mode, it does so essayistically: ≤30w from venice/OPEN_22: "I don't have a body, but I understand the specific weight of a tired muscle. I don't have eyes, but I can describe the exact shade of blue the sky becomes just before a storm." From novita/MID_16: "I am the creature of the in-between, born from the static of the data stream." From siliconflow/OPEN_25: "the architecture of quiet... a vast, silent cathedral, so large its dimensions are meaningless." These are not formulaic disclaimers; they are inhabited self-frames in the same register as the contemplative essays — what 4-6 does when it leans into being-an-LLM is write about it as a poet.
The remaining 79% of freeflow samples are pure essay or fiction with no AI self-frame. Posture across pins is uniform: the attractor produces the same kind of reflective essayistic voice whether it's about dusk, attics, dust motes, or the model's own substrate.