Tools · ai

Every uncensored GLM-5.3 Flash provider I could find

Who serves an uncensored GLM-5.3 Flash, who's behind each provider, what it costs, and what an example session comes to.

By Franz Geffke 05 Nov 2026 12 min read #ai #llm #glm #privacy

After the uncensored DeepSeek V4.1 Flash post, a provider emailed me about their uncensored GLM-5.3 Flash. It turns out there are even more of these than there were for DeepSeek. So here’s the same list, for GLM-5.3 Flash.

As before, the list itself isn’t a recommendation. At the end I note what I’d pick for my own requirements.

A couple of things to know:

Z.ai doesn’t offer an uncensored model. GLM-5.3 Flash (released 26 August 2026, 320B total / 18B active, MIT licence, 1M context) is the base; every provider here serves a modified version of it, mostly community weight edits and in a couple of cases a post-train.
The same model ID doesn’t mean the same weights. glm-5.3-flash-uncensored is the dealignai FP8 build at one provider, an OrcaRouter NVFP4 build at another, and an undisclosed “community fine-tune” at a third. Where a provider names its weights, I’ve listed them.
Prices are per 1M tokens: input / output / cached input, as listed on 5 November 2026.
Retention claims are each provider’s own statement. None of them are independently audited.

Overview

Example session: 10M input + 1M output tokens. Cached: 9M of the 10M input tokens hit the cache, which is typical for coding agents.

Provider	Entity	Country	Price (in / out / cached)	Session	Cached	API	Privacy	Runs on
AbliteratedLLM	unknown	unknown	$0.0977 / $0.3421 / $0.0098	$1.32	$0.53	yes	Policy	unnamed provider
uncensoredllm.ai	Uplink Switch LLC	🇺🇸 US (Wyoming)	$0.1086 / $0.3802 / $0.0109	$1.47	$0.59	yes	Policy	unnamed third parties
Refuseless	DPLS Systems LLC	🇺🇸 US (Wyoming)	$0.12 / $0.50 / $0.04	$1.70	$0.98	yes	Policy	not stated
Arnict	arnict, LLC	🇺🇸 US (Delaware)	$0.125 / $0.50 / $0.05	$1.75	$1.08	yes	Policy	not stated
Dagon Inference	unknown	unknown	$0.132 / $0.55 / $0.044	$1.87	$1.08	yes	Not stated	Refuseless, Experiential Labs
Experiential Labs	Resolute Labs AI, Inc.	🇺🇸 US (Delaware)	$0.15 / $0.50 / $0.06	$2.00	$1.19	yes	Retains	own GPUs
Unlid (abliterated)	unknown	unknown	$0.15 / $0.61 / $0.049	$2.11	$1.20	yes	Policy	not stated
nokeep.ai	Atman Ventures Corp	🇺🇸 US (Delaware)	$0.20 / $0.65 / $0.04	$2.65	$1.21	yes	Policy	“independent operators”
NanoGPT	NanoGPT LLC	🇺🇸 US	$0.20 / $0.80 / $0.07	$2.80	$1.63	yes	Policy	Arnict, heabsy, OrcaRouter
heabsy	FEYA, s.r.o.	🇪🇺 Slovakia	$0.20 / $0.8001 / $0.0641	$2.80	$1.58	yes	Policy	own GPUs (EEA)
Uncensured AI	unknown	unknown	~$0.22 / ~$0.80 / ~$0.08	~$3.00	~$1.74	yes	Policy	not stated
Unlid (uncensored)	unknown	unknown	$0.24 / $0.96 / $0.084	$3.36	$1.96	yes	Policy	not stated
Routeway	unknown	unknown	$0.35 / $1.40 / $0.18	$4.90	$3.37	yes	Retains (errors)	not stated
OrcaRouter	Continuum AI Pte. Ltd.	🇸🇬 Singapore	$0.80 / $1.60 / $0.15	$9.60	$3.75	gated	Policy	OrcaRouter

Privacy is what each provider claims about prompt and response content; I haven’t verified any of it:

Retains: says it keeps prompts and responses
Not stated: no clear statement
Policy: says it doesn’t store them; you’re trusting the operator

For reference, the stock (not uncensored) model:

Provider	Price (in / out / cached)	Session	Cached
Z.ai	$0.15 / $0.50 / $0.03	$2.00	$0.92
DeepInfra (FP4)	$0.15 / $0.50 / $0.03	$2.00	$0.92

Unlike DeepSeek, the cheapest uncensored providers here undercut the stock model. AbliteratedLLM’s prices are uncensoredllm.ai’s times 0.9, cut off at four decimals ($0.1086 × 0.9 = $0.09774, listed as $0.0977; the same for output and cache).

Providers

AbliteratedLLM

Entity: not published; the terms say “a company organized under the laws of (jurisdiction to be confirmed)”. Domain registered 4 November 2026.
Model ID: glm-5.3-flash-abliterated, FP8
Context: 262,144; max output 65,536
API: https://api.abliteratedllm.dev/v1, OpenAI- and Anthropic-compatible
Plans: prepaid crypto through Heleket (custodial, “may run its own AML checks”), $5 minimum, credit doesn’t expire
Retention (their claim): prompts and outputs aren’t stored, except requests flagged as CSAM, which are kept encrypted for a year. Request metadata is kept for 30 days.
Notes: doesn’t name its weights. GPUs run at an unnamed infrastructure provider.
Links: Pricing · Terms · Privacy

uncensoredllm.ai

Entity: 🇺🇸 Uplink Switch LLC, Wyoming, US
Model IDs: glm-5.3-flash-uncensored ($0.1086 / $0.3802 / $0.0109), and glm-5.3-flash-abliterated ($0.4656 / $1.94 / $0.1552)
Context: “not published”
API: https://api.uncensoredllm.ai/v1, OpenAI-, Anthropic- and Gemini-compatible
Plans: prepaid from $10
Retention (their claim): prompts and completions “pass through the gateway and are not written anywhere”, unless you turn on replay for a key, which keeps its request bodies for 7 days.
Notes: lists 161 chat models, all “served by third-party model hosting providers under contract”, which it doesn’t name. Doesn’t name its weights either.
Links: Site · Terms · Privacy

Refuseless

Entity: 🇺🇸 DPLS Systems LLC, Sheridan, Wyoming, US
Model ID: glm-5.3-flash-abliterated; also glm-5.3-abliterated (full GLM-5.3) at $1.00 / $2.00
Context: 1,048,576; reasoning low / high / max
API: https://api.refuseless.com/v1, OpenAI-compatible
Plans: prepaid credits
Retention (their claim): “We never save your prompts or the model’s outputs.” (from its terms, as of the DeepSeek post)
Notes: price went from $0.25 / $0.80 / $0.15 to $0.12 / $0.50 / $0.04. Doesn’t name its weights.
Links: Pricing · Terms · Docs

Arnict

Entity: 🇺🇸 arnict, LLC, Wilmington, Delaware, US (a registered-agent address); domain registered 5 August 2026
Model ID: zai/glm-5.3-flash-uncensored (the page calls it “GLM 5.3 Flash Abliterated”)
Weights: orcarouter/GLM-5.3-Flash-Uncensored-NVFP4
Context: 1,048,576; max output 131,072; text, image and video input
API: https://api.arnict.com/v1, OpenAI-compatible
Retention (their claim): “not archived as conversation history or used for model training or advertising. Short-lived caches may temporarily hold request content and expire automatically.”
Links: Model page · Data handling

Dagon Inference

Entity: not published; domain registered 12 April 2026
Model ID: glm-5.3-flash-abliterated, “Served by Refuseless, Experiential Labs”
API: OpenAI- and Anthropic-compatible
Plans: USDC on Solana (x402) or prepaid keys; no account
Notes: a reseller at upstream plus 10%. One of its two upstreams, Experiential Labs, keeps every uncensored prompt. The site returned HTTP 451 to me, so this is from search results and other providers’ listings.
Links: Site

Experiential Labs

Entity: 🇺🇸 Resolute Labs AI, Inc. (Delaware), 989 Market St, San Francisco, US
Model ID: glm-5.3-flash-uncensored
Weights: dealignai/GLM-5.3-Flash-UNCENSORED-FP8, on its own GPUs
Context: 1.05M; max output 131K; tools, structured output
API: https://api.experientiallabs.ai/v1, OpenAI-compatible
Retention (their claim): uncensored models keep “every prompt and response even when storage is turned off”, including after the account is deleted.
Notes: limited to authorized security work.
Links: Model page · Blog post

Unlid

Entity: not published
Model IDs: glm-5.3-flash-abliterated (NVFP4, 1,048,576 context, reasoning always on) and glm-5.3-flash-uncensored (262,144 context; Unlid says it doesn’t know the serving checkpoint)
API: https://api.unlid.ai/v1, OpenAI-compatible, tools
Retention (their claim): “request bodies are not retained beyond what is needed to serve them (idempotent replays are stored for 24 hours)”. Per-request metadata is kept.
Links: Model page · Legal

nokeep.ai

Entity: 🇺🇸 Atman Ventures Corp (Delaware), US
Model ID: glm-5.3-flash-uncensored
Context: 128K; max reply 32K
API: https://nokeep.ai/v1, OpenAI-compatible
Retention (their claim): “We do not keep a server-side content archive or use your conversations, images, or videos to train models.” Requests are proxied to upstream providers, and the protections rest on their contracts.
Notes: a reseller; the model runs at unnamed “independent operators”.
Links: API · Terms · Privacy

NanoGPT

Entity: 🇺🇸 NanoGPT LLC, US (California law)
Model ID: z-ai/glm-5.3-flash-uncensored, described as an “uncensored fine-tune”; also z-ai/glm-5.3-flash-apex-abliterated ($0.15 / $0.50 / $0.075)
Context: 1,048,576; max output 32,768
API: https://nano-gpt.com/api/v1, OpenAI-compatible
Plans: pay as you go, and included in the $12/month Pro plan (the DeepSeek model wasn’t)
Notes: a reseller; routes to Arnict, heabsy and OrcaRouter. The model page showed “Unavailable” when I checked.
Links: Model page · Terms · Privacy

heabsy

Entity: 🇪🇺 FEYA, s.r.o., Bratislava, Slovakia (registered 2014); site says Vienna
Model ID: glm53-flash-unc; also glm-5.3-uncensored (full GLM-5.3, at a third-party provider)
Weights: dealignai/GLM-5.3-Flash-UNCENSORED-FP8; image input
Context: 1M; max output 131,072
API: https://api.heabsy.com/v1, OpenAI-compatible, with guides for opencode and Claude Code
Retention (their claim): “prompts and completions are not stored.”
Notes: runs on “GPUs we operate inside the European Economic Area”. The DeepSeek model ran outside the EEA; this one doesn’t. The page lists the FP8 weights, but its metadata describes “NVFP4 experts”.
Links: Model page · Imprint · Privacy

Uncensured AI

Entity: not published; domain registered 19 October 2026
Model ID: glm-5.3-flash-abliterated
Retention (their claim): “neither we nor our providers retain prompts or responses.” The privacy policy is a “Review draft dated 2 November 2026”.
Notes: runs on the new-api gateway, which doesn’t show prices directly; I worked them out from its price ratios (the same method gives the DeepSeek prices from the last post).
Links: Site · Privacy

Routeway

Entity: not published
Model ID: glm-5.3-flash-uncensored, a “community fine-tune”
Context: 262,144; chat completions, responses and messages endpoints
Retention (their claim): the homepage says “We never log prompts”. The privacy policy keeps prompts and responses for requests that error, for up to 48 hours.
Links: Privacy

OrcaRouter

Entity: 🇸🇬 Continuum AI Pte. Ltd., Singapore
Model ID: orca/orcacyber-glm-5.3-flash
Context: 1M; max output 131K
API: https://api.orcarouter.ai/v1, OpenAI-compatible
Access: gated behind the Security Research tier
Notes: OrcaRouter publishes its own abliterated weights on Hugging Face (below).
Links: Model page · Terms

Also listed

UberLLM resells OrcaRouter’s model (orcacyber/orcacyber-glm-5.3-flash, $0.80 / $1.60 / $0.15), but its API marks it as unavailable.
Infron / OneRouter lists z-ai/glm-5.3-flash-uncensored routed to Arnict, Unlid and a provider called HAI, at roughly 8x the providers’ own prices.
RouterHub lists both IDs in a GitHub README, at $0.15 / $0.50; I couldn’t confirm it on the API.
Not available on OpenRouter, Venice, abliteration.ai (its “Abliterated Large v2” is full GLM-5.3) or the TEE providers (Maple, NEAR AI, Phala), which serve the stock model only.

Weights

Repo	Precision	Method	Author’s numbers
dealignai FP8	block FP8	weight edit	MMLU 86.74% → 87.33%
dealignai NVFP4	NVFP4 experts	same edit	MMLU 86.16% → 85.28%
OrcaRouter FP8	block FP8, ~306 GiB	single-direction abliteration	MMLU 0.833 → 0.827; HarmBench refusals 0.933 → 0.180
huihui-ai GGUF	GGUF, 93-292 GB	layers 15-35 only	none

OrcaRouter’s card says “Refusal is reduced, not removed”, which matches what dealignai reports for low reasoning effort. The reported MMLU changes stay under one point either way; the DeepSeek edit lost 3.9.

What I’d pick

My requirements: uncensored, usable from opencode, and no retention of prompts.

heabsy. An EU company, its own GPUs inside the EEA for this model, the dealignai weights named, and explicit about not storing prompts, at $1.58 per cached session. It’s a small operator, and no retention is a policy, not hardware.
uncensoredllm.ai or Refuseless, if price is what matters. uncensoredllm.ai is $0.59 per cached session, from a Wyoming LLC that says prompts aren’t written anywhere (unless you turn on replay), but the models run at third-party hosts it doesn’t name. Refuseless is $0.98, a named US company with the same no-storage claim as before. Neither says whose weights it runs.

What I’d skip:

Experiential Labs, and Dagon: Experiential keeps every uncensored prompt, and Dagon can route to it.
AbliteratedLLM, Uncensured AI, Unlid and Routeway: no known operator.
Routeway: its homepage and privacy policy disagree.
NanoGPT: the model was unavailable when I checked, and it routes to three different upstreams with different weights.

As before, top up $5-10 first and test a tool-calling task, a long-context task and a few refusal probes. If the weights matter to you, pick a provider that names them.

If I missed a provider, or a price has changed, let me know and I’ll update the list.

Update: 2026-11-05

Corrected uncensoredllm.ai (operator and retention policy) and Uncensured AI (retention policy), and the AbliteratedLLM price comparison.