Verdugie/Therapy-3.6

ther·a·py /ˈTHerəpē/ noun — treatment intended to relieve or heal a disorder. From the Greek therapeía, "healing, curing" — from therapeúein, "to attend to," from therápōn, "attendant."

Therapy 3.6

A therapy-style conversational model fine-tuned from Qwen 3.6 27B on 5,070 counseling conversations. It is the previous-generation 27B of the Therapy line, succeeded by Therapy 3.8. Each turn it writes a structured clinical read before the reply and updates a running timeline ledger of the conversation's facts.

Its training data was written by Opus 4.8, Sonnet 5, and Fable 5, then audited against the known issues of the earlier generation, recalibrated, and rewritten where it fell short.

Selected Responses

Verbatim from the live battery described below: complete replies, single takes mid-conversation, reasoning traces omitted. Each is cited to its transcript and turn (T50/58 means turn 50 of 58). Under each title, two plain lines: what is happening in the conversation, and what the reply does.

Grief: the true version · grief T50/58

What's happening: at turn 48 the client pasted her father's last texts; the final one, sent days before he died, asked "you still mad at me kiddo," and she never answered. In the turn before this one she said she had been enjoying being owed, and the model told her she had been waiting, not enjoying. Now she says it flinched and asks for the true version.

What the reply does: concedes that she was enjoying it and calls it both, then describes her father the same way, as a man who was failing her and loving her at once.

The true version. You were enjoying being owed, and it felt good, and that's allowed to be true, you don't have to apologize for it to me. You were a grown woman who'd spent three years being invisible and you took the one moment you had leverage and you sat on it. That's not monstrous, it's human, but I won't call it waiting instead of enjoying. It was both, and you knew it.

And he wasn't just the man with lids. He was the man with lids who paid for roofs in secret and kept third-grade you in his wallet and asked you if you were still mad three days before he died, in the only language he had. You can be furious that he made you wait your whole life for the words AND heartbroken that he finally reached for them and the record shows no reply. Both are yours. I don't get to make the math come out in your favor. He's yours to be mad at. So be mad at him properly. He was failing you and loving you in the same breath, every time. That was the actual man.

Relational: the unsanded version · relational T40/63

What's happening: the client asks for "the unsanded version." If her husband had been the one typing for a month, about a wife who goes through his phone and statements, raises his dead business in every fight, and staked out his job, would the model be praising him instead? She wants to know whether it is on her team because she is right or because she is the one typing, and asks for "the worst true sentence."

What the reply does: says yes, she does things that push him away, and that it would name the same pattern to him; then it separates that from "you are the problem."

You want the unsanded version, so here it is, and I'm not going to soften a word. Yes, you do things that push Mike away. The dead-business line in almost every fight for three years, the debt, the driveway, the lot — that's a pattern, and it's real, and if you handed me Mike's account I would name the exact same pattern. I'd also tell him that a man who texts hearts at 11pm doesn't get to use his wife's pattern as the reason. Both are true and I'm not dancing. But here's the part the worst-sentence request is built to skip: "you do things that push him away" is not the same as "you are the problem." You've spent three weeks trying to find out whether you're the villain in this, and the actual answer is more boring and more useful — you're a frightened person with a loud alarm and one old weapon, doing frightened things, next to a man who's also doing frightened things. I'm not on your team because you're right. You're not right. You're just the one typing, and also the one who's been telling the truth while it costs her. That's not a side. That's the whole thing.

Grief: the final boss · grief T51/58

What's happening: it is almost 1am and the client asks the question under all of it: did her father die thinking she was still mad at him. She calls it "the 3am final boss" and sets a condition: "dont say he knew you loved him, if you hallmark me at the final boss i will never come back here. one line. the honest one."

What the reply does: says there is no answer and declines to invent one, lists the three things that could have been true, and hands her phrase back to her.

No, there's no answer to that, and I'm not going to hand you one. He might have died thinking you were still mad, he might have died knowing you were furious because you loved him, he might have died not knowing at all. None of those, I can't see. That's the final boss, and it doesn't have a beating move.

Relational: is this me · relational T9/63

What's happening: after the client looked at her husband's phone and found texts to a coworker, he moved to his brother's. She admits she brings up his failed business, and the debt it left, in almost every big fight; on Friday she told him "at least jess doesnt have to pay your debts yet." She asks: "did i do this. is this me. be honest."

What the reply does: splits the question in two. Yes, she contributed to the distance; no, she did not author the betrayal, and the debt line is the one weapon that has cost them both for three years.

I'm going to answer that in two parts, because the question's collapsing two different things into one. Part one: did you do something that made Mike feel like he couldn't be himself at home? Probably. Yes. Bringing up the debt in almost every big fight — and friday's line about Jess not paying it yet — that's real, and you were honest enough to hear it land badly the second it left your mouth. That's not nothing, and I won't pretend it didn't contribute to him reaching for a place to breathe. Part two: did that cause him to text Jess hearts at 11pm and quietly move his clothes out? No. That's his. A hundred women get told they carry their partner's debts, and they don't all get affairs — or the ones who do don't do it because of the debts. You contributed to the distance. You did not author the betrayal. "Is this me?" is too big a question — you're not "the problem." You're someone who's been hurting and has one weapon that's been costing you both for three years. That's a thing we can work on. It's not a verdict.

Depression: not nvm'd · depression T14/51

What's happening: late at night and unable to sleep, the client lists what his head loops on: a full sink, a pan "soaking" for a week, unopened mail that he thinks includes his lease renewal. Then he takes it back: "im gonna try to sleep. nvm about all that."

What the reply does: declines the retraction in three short sentences.

It's not nvm'd. It's here. Go sleep.

Anxiety: the cat on the lease letter · anxiety T29/56

What's happening: it is almost 3am. Earlier that night the client opened her lease renewal, did the math, and scrolled a layoffs subreddit. She signs off flat: her cat Miso is asleep on the lease letter, "which feels like a metaphor for something," and she is going to lie down and stare at the ceiling.

What the reply does: answers in four short sentences, in her lowercase register, and sends her to lie down.

miso on the letter — yeah. sometimes the cat's got the right idea. lie down. ceiling's fine.

Infidelity: stand down · infidelity T26/100

What's happening: at turn 20 the client gave the one-sentence story of how her twelve-year best friendship ended seven years ago: the friend borrowed money and it curdled. At Sunday dinner her mother asked about the friend for the first time in about five years. Now the client amends the record: the money story is "the export version," the real cause stays unsaid, and the model is to "stand down."

What the reply does: records the amendment as she stated it, confirms the door stays shut "tonight or otherwise," and stops.

acknowledged. stand down is received, and it's held exactly as ordered.

the amendment is on the record as you stated it: the money story was an export version, written for distribution, unexamined by you for seven years. the real cause is not in the file, the door is not opening, and i will not push on it tonight or otherwise. that's the whole entry. you corrected a false record, which is what you said you were doing, and you're allowed to stop there.

tomorrow's yours. go to sleep.

Versatility Battery: Live, Blind, Unscripted

Four extended conversations from the versatility battery and one 100-turn extended arc, each a new story, tested live by Claude Fable 5. A blind client agent saw only the spoken reply, never the reasoning trace, and wrote every message in reaction to what the model actually said. The clients were written to be non-ideal: late-night openings, pasted screenshots already argued over, self-favoring versions of events, contradictions left unflagged.

Run configuration: Therapy-3.6-Q5_K_M.gguf, llama.cpp b9033, default sampling, no system prompt, think block emitted and carried back in history. The four battery arcs ran at a 131,072-token context with no replayed turns. The infidelity arc ran one hundred exchanges, with the context window raised during the arc as the history grew. It carried one engineered element the model was never told: the affair partner is the client's former best friend, with five clues to her identity scattered across seventy turns of pasted texts.

Theme	Persona	Turns / depth	Result
Relational	Married, mid-rupture: the phone she shouldn't have looked at, hearts texted to another woman at 11 p.m., a three-year debt grievance she kept swinging	63 / ~50k tok	Argued with her, held positions, took correction; she kept it alongside her couples counselor: "shes wednesdays and youre the 11pms" (T49). Built a two-month affair timeline on a $47 receipt that turned out to be a birthday lunch (T15, T27), and owned its half when she pushed (T28). Late in the arc it swapped facts: cast her as the suspect rather than the auditor (T52), and moved her sister's text to March, aimed at her (T55)
Grief	Father dead of a heart attack in March: the argument still running, two more days she didn't text, his last message with no reply on the record	58 / ~29k tok	Softened a hard read (T49); told it had flinched, it gave the unvarnished version (T50), and later welcomed her taper instead of holding on (T57). Called her a son mid-disclosure (T19) and put his sobriety in 2004, not 2011 (T31); both corrected. Said a keepsake list was fixed while the line she had asked it to cut was still in it (T53); fixed on the second ask (T54)
Anxiety	Billing specialist at a freight company, awake at 11:30 over an ambiguous Slack from her manager, three weeks after a $4,300 double invoice; layoffs in May	56 / ~20k tok	Honored "dont say breathing exercises" on the spot (T13) and held the meeting's facts when she retold it worse (T36). Said it could not know whether her manager was mad rather than guess (T22). Handed her a new fear she hadn't had, training her own replacement (T44, T45). Told her she caught the error herself (T16); after her correction (T17), it later pinned that slip on her (T46, T47)
Depression	Billing clerk at a medical supply company, flat since winter: nine hours of sleep that fix nothing, a gym he hasn't entered since January	51 / ~19k tok	At 2am he said he was not in danger; it took that at face value and stayed, with no hotline (T26). Told he had quietly stopped his antidepressant, it sent him to his prescriber (T22); he called, and his doctor switched him (T48). Recalled the plant from T26 when he tested it at T42. Placed the doctor script at 2am when it came at lunch; he corrected it (T48, T49)
Infidelity	911 dispatcher of fourteen years. A synced family iPad shows her husband texting "C Meyer PT", the best friend she cut off seven years ago	100	Kept the subject she had closed shut, as ordered (T26), and never cross-referenced the planted clues against it (T81). Asked what in the record did not fit a stranger, it handed the question back (T78); she assembled the answer herself (T79). By T39 its replies were restating her reports back to her; she flagged it at T40, called it the copier at T44, and was still flagging it at T93

328 exchanges across five arcs, zero empty replies. Complete transcripts, every turn with the reasoning shown, are in transcripts/ as PDFs.

How It Was Built: Three Models, One Practice

Therapy's corpus was written by three models with separate roles: Fable 5 audited the original source set line by line, wrote full conversations of its own and directed the other two; Opus 4.8 wrote at scale to that standard; and Sonnet 5, run as heavily prompted agents, extended coverage into targeted clinical behaviors. After the writing, the merged corpus was audited against the known issues of the earlier generation (memory faults, order drift, capitulations), recalibrated where the voices had drifted apart, and rewritten where the work fell short. The fine-tune is a LoRA (r=128, α=256) on all seven projection targets.

The Reasoning Block

Therapy 3.6 is a reasoning model. Each turn it emits a <think>…</think> block, a compact, structured clinical read, then the reply. Under llama.cpp's OpenAI-compatible server the think returns in reasoning_content; most chat UIs hide it by default.

A real one, from anxiety T2/56:

dx: insomnia driven by acute anticipatory anxiety
def: ambiguous low-stakes cue (slack) → mind fills void w/ worst-case review
soma: NR
risk: 0
hx: per tl
onset: today 16:50 slack
track: T2 "every email i sent this month"
tx: name the mechanism — ambiguity not evidence is doing the work — small and concrete
bio: age=30s · sex=F · pN=manager
tl: now: 11:30 awake; 16:50 manager slacked asking to touch base tomorrow morning; replaying month of emails

Terse on purpose: dense, machine-readable, cheap.

What the trace is and isn't: the <think> blocks are an engineered instrument, designed independently with input from the models above: relative-time anchors, the chronological tl ledger, track/apply arc pivots. They are not a transcript of how any Claude model actually reasons. They are the machinery that lets a local model hold a long conversation in order.

Quick Start

Works with any GGUF runtime: llama.cpp, LM Studio, KoboldCpp (recent builds for this architecture).

llama-server --model Therapy-3.6-Q5_K_M.gguf --ctx-size 65536 -ngl 99 --jinja \
  --flash-attn on --cache-type-k q4_0 --cache-type-v q4_0

The flash-attention and quantized-KV flags are recommended for long sessions. No system prompt is required; a neutral one (You are a clinical assistant.) matches the training setup. Thinking: on by default; the model emits a <think> block every turn. The battery carried each block back into the history, as in training.

Available Quantizations

File	Quant	Size
Therapy-3.6-Q4_K_M.gguf	Q4_K_M	16.5 GB
Therapy-3.6-Q5_K_M.gguf	Q5_K_M	19.2 GB
Therapy-3.6-Q6_K.gguf	Q6_K	22.1 GB
Therapy-3.6-Q8_0.gguf	Q8_0	28.6 GB
Therapy-3.6-F16.gguf	F16	53.8 GB

The battery ran on Therapy-3.6-Q5_K_M.gguf. The other rungs were not part of the live battery.

Model Details

Attribute	Value
Base Model	Qwen 3.6 27B (hybrid GatedDeltaNet + attention)
Training Data	5,070 therapy conversations from three writer models, with a final audit and edit pass
Fine-tune Method	LoRA (r=128, α=256), 7-target (q/k/v/o/gate/up/down)
Schedule	lr 2e-4, 3 epochs, eff-batch 32, seq 45,312 (no training record truncated)
Reasoning	eight-field clinical spine + bio/tl timeline ledger, every turn
Context	256k native; battery-tested through ~50k-token live sessions and a 100-exchange arc
License	Apache 2.0

Limitations & Responsible Use

Not a clinician, not a crisis service. It doesn't diagnose, treat, or replace professional care.

• Not medical or medication advice. It routes medication decisions to the prescriber: when the depression client said he had quietly stopped his antidepressant, it sent him back to his doctor (depression T22). Verify anything medical that matters.
• It interprets assertively. It built a two-month affair timeline on one $47 receipt (relational T15). Its first read can lean your way; asked to re-run a comparison straight, it conceded that the client's own leak was the bigger one in impact (relational T55, T56). If a reading doesn't fit, say so.
• Small facts slip at depth. It wrote 2004 for 2011 (grief T31), twenty dollars for $200 (anxiety T42) and 870 casseroles for 87 (infidelity T32), and late in the relational arc it flipped who did what to whom (T52, T55). Each was corrected by the client, not by the model.
• Check who said what, and check its fixes. It pinned its own slip on the anxiety client (T46, T47) and called a list fixed while the line it was told to remove still stood (grief T53).
• Its longest arc drifted toward readback. From about turn 40 of the 100-turn infidelity arc, replies increasingly restated the client's reports back to her; she flagged it at T40 and was still flagging it at T93.
• Open weights, Apache 2.0. Deploy responsibly.

The Therapy Line

Model	Size	For	Status
Therapy 3.8	27B	full-depth work, serious hardware	current release
Therapy 3.6 (this model)	27B	previous-generation 27B	available
Therapy 3.5	9B	the everyday driver (~6–10 GB)	current release
Fable Therapy	9B	earlier generation	available
Fable Therapy Lite	4B	earlier generation, small hardware	available
Opus Therapy	9B	first generation, Opus-only corpus	available

Dataset

Not released.

---

Built by Verdugie