At a Glance #

Claim IDs resolve in sources.md and the claims ledger. Benchmark rows are descriptive unless their declared comparison group and design support comparison; unknown is not zero. See the report and full CSVs for evaluation conditions, policy exceptions, and cost assumptions. Wide tables scroll horizontally on the site.

Model and harness setups #

id	model	harness	configuration	task context	availability	strengths	limitations	claim ids
astra	gpt-6-astra	Codex CLI/IDE local	Start high; escalate effort only after checking failure	Complex implementation and debugging	API paid tiers; client access depends on account	Steerable long-running candidate	No model-only superiority established; costly cache/long requests	CR01;CR02;CH17;CH20
fable51	claude-fable-5-1	Claude Code	High default; preserve prefixes and verify fallback route	Hard refactoring and long sessions	GA; Pro/Max/Team/Enterprise and listed clouds	Cache-heavy escalation candidate	Retention eligibility; API breaking changes; human review needed	CR03;CR04;CR05;CR06;CH05;CH06
opus5	claude-opus-5	Claude Code	High; bounded task and tests	Everyday implementation	Available since July 24	Lower list rates than Fable; vendor recommended starting model	Vendor guidance; no measured repo-specific preference	CR07;CR16
sol	gpt-5.6-sol	Codex	Medium starting effort; bounded task	Everyday implementation	Listed active model; account access varies	Lower current token rates	Promotion; no accepted-task savings established	CR13;CH17
gemini38	gemini-3.8-flash	Antigravity managed agent	Medium default; allow billed reasoning	Greenfield/UI and bounded budget evaluation	GA September 2	Native managed route; managed-agent economics unmeasured	May spend more tokens; no independent advantage established here	CR08;CR09
deepseek4	deepseek-v4-pro	Codex	Vendor Responses configuration; high starting effort	Budget workflow compatibility evaluation	GA hosted Pro0813; Flash0731 separately	Documented Responses integration	Not tested; Responses protocol has unsupported fields and tools	CR11;CR12;CR18
glm53	glm-5.3	OpenCode	Reasoning enabled; explicitly select effort; retain task tests	Budget-conscious implementation and debugging with human review	All Coding Plan tiers; API account restrictions documented	Low posted token rates; documented Coding Plan support	Task acceptance cost unmeasured; aliases reroute; account protocol restrictions; Coding Plan privacy mapping unresolved	CGL01;CGL04;CGL06;CGL20;CGLP03
glm53eval	glm-5.3	Claude Code	max effort; 400K context; 128K output; temperature=1; top_p=1; Tool Search disabled	Vendor Terminal-Bench 3.0 evaluation	Vendor reported result	Detailed vendor evaluation conditions	No independent reproduction or matched deployment measurements	CGL02
glm53flash	glm-5.3-flash	OpenCode	Explicit effort; max default; multimodal model	Low-price, bounded implementation tasks to pilot	All Coding Plan tiers; open weights available	Lower durable list rates; multimodal input; MIT weights	Quality and accepted-task savings unverified; not a small local model	CGL05;CGL06;CGL08
local8	Qwen/Qwen3.5-4B	Qwen-Agent + vLLM/SGLang	Proposed 4-bit weights; KV precision and context to measure	8 GB dedicated VRAM capacity-screening pilot	Released weights or documented service through cutoff	Small Apache 2.0 candidate with documented local agent route	Quant format, runtime memory and task reliability not tested	CGL17;CGL18;CGL21
local16	Qwen/Qwen3.5-9B	Qwen-Agent + vLLM/SGLang	Proposed 8-bit weights; KV precision and context to measure	16 GB dedicated VRAM capacity-screening pilot	Released weights or documented service through cutoff	Larger small-model candidate with local MCP route	No measured agentic reliability or latency	CGL17;CGL18;CGL21
local24	Qwen/Qwen3.8-27B	vLLM/SGLang endpoint + chosen agent	Proposed 4-bit weights; preserve reasoning per model docs; client verification required	24 GB dedicated VRAM local pilot	Released weights or documented service through cutoff	New dense model; Apache 2.0 designation; vendor coding improvement	Vendor benchmark does not validate this 4-bit setup; client interoperability untested	CGL09;CGL10;CGL18;CGL21
localbig	Qwen/Qwen3.8-27B	Selected local serving stack and agent	Proposed 8-bit weights; engine/hardware support to verify	64 GB unified-memory capacity-screening pilot	Released weights or documented service through cutoff	More weight precision headroom than the 24 GB tier	Unified memory shared with OS; neither engine compatibility nor latency measured	CGL09;CGL18;CGL21
qwen38eval	Qwen/Qwen3.8-27B	Terminus	Vendor benchmark; detailed configuration unknown	Vendor Terminal-Bench 2.1 result	Vendor reported result	Reported score of 73.0	No 4-bit or commodity GPU inference established	CGL10
kimik3	kimi-k3	Kimi Code CLI	Preserve full assistant messages; effort low/high/max; max default	Long-session engineering candidate	Documented API and released weights	Native multimodality and 1M context documented	Vendor capability claims; custom license; huge self-hosting memory	CGL13;CGL14;CGL18
minimaxm3	MiniMax-M3	Selected compatible agent	Choose enabled/adaptive/disabled thinking; integration test required	Multimodal coding candidate	Released weights or documented service through cutoff	Documented 1M context and serving routes	Custom license; no measured task economics or client validation in this review	CGL15;CGL16
qwen38next	Qwen/Qwen3.8-Flash-Next	Selected compatible agent	Experimental preview; version/quantization and parser verification required	Larger-server experiment	Open-weight experimental architecture preview	Sparse architecture with offloadable embeddings	Custom AI Work Assistant/MaaS licensing; 6B active parameters do not determine total weight memory	CGL11;CGL12
TO1	GPT-6 Astra	Codex	Max effort; Harbor evaluation using Terminal-Bench 4.0.0; remaining settings unknown	66 terminal-work tasks in Terminal-Bench 4.0.0	Observed benchmark submission; current product access documented separately	Published terminal-task resolution result	Exact harness version, fallback routing, complete budget and evaluation date unknown; descriptive result only	CO04;CO03
TO2	Claude Fable 5.1	Claude Code	Max effort; Harbor evaluation using Terminal-Bench 4.0.0; remaining settings unknown	66 terminal-work tasks in Terminal-Bench 4.0.0	Observed benchmark submission; current product access documented separately	Published terminal-task resolution result	Exact harness version, fallback routing, complete budget and evaluation date unknown; descriptive result only	CO05;CO03
TO3	GLM-5.3	Claude Code	Max effort; Harbor evaluation using Terminal-Bench 4.0.0; remaining settings unknown	66 terminal-work tasks in Terminal-Bench 4.0.0	Observed benchmark submission; current product access documented separately	Published terminal-task resolution result	Exact harness version, fallback routing, complete budget and evaluation date unknown; descriptive result only	CO06;CO03
TO4	Claude Fable 5.1	Claude Code	Endor Labs Agent Security League evaluation; exact inference settings unknown	Python tasks derived from CVE patches	Benchmark row dated 2026-09-04	Functional and security-test outcomes reported separately	Exact September run settings, adjusted denominator and uncertainty unknown	CO07
copilot-team	gpt-6-astra	GitHub Copilot app/CLI	Business/Enterprise policy; exclusions and user budgets; human review	Existing GitHub team with approved hosted-data route	GA rollout gradual; admin model enablement	Central policy and credit controls; documented Astra route	Provider ZDR does not establish GitHub service retention; task advantage unmeasured	CH02;CH04;CH07;CH12;CH21

Benchmark observations #

setup id	benchmark	benchmark version	evaluator	evidence type	comparison group	comparison basis	metric	score	unit	uncertainty	claim ids
glm53eval	Terminal-Bench	3.0	Z.ai	vendor-experiment	none	none	avg@3 official verifier score	28.3	reported score (0–100)	unknown	CGL02
qwen38eval	Terminal-Bench	2.1	Qwen	vendor-experiment	none	none	Vendor reported score	73.0	reported score (0–100)	unknown	CGL10
TO1	Terminal-Bench	4.0.0	Terminal-Bench/Harbor; contributor presentation by Snorkel AI	independent-experiment	none	none	Resolution rate	58.2	percent	Reported ±2.8 percentage-point margin reflecting run-to-run variance; interval level unknown	CO04;CO03
TO2	Terminal-Bench	4.0.0	Terminal-Bench/Harbor; contributor presentation by Snorkel AI	independent-experiment	none	none	Resolution rate	57.9	percent	Reported ±3.8 percentage-point margin reflecting run-to-run variance; interval level unknown	CO05;CO03
TO3	Terminal-Bench	4.0.0	Terminal-Bench/Harbor; contributor presentation by Snorkel AI	independent-experiment	none	none	Resolution rate	41.8	percent	Reported ±3.2 percentage-point margin reflecting run-to-run variance; interval level unknown	CO06;CO03
TO4	Agent Security League / SusVibes	unknown	Endor Labs	independent-experiment	none	none	FuncPass	87.2	percent	unknown	CO07
TO4	Agent Security League / SusVibes	unknown	Endor Labs	independent-experiment	none	none	SecPass	37.4	percent	unknown	CO07

If you are X, use Y #

audience	scenario	prerequisites	pick	alternative	tradeoffs	confidence	confidence reason	would change if	claim ids
Individual	Daily implementation	Tests; review	Keep Sol/Codex; reassess	Opus 5/Claude Code	Promotional rates; task costs unmeasured	medium	Documented rates; limited experience	Another setup reduces accepted-task cost	CR13;CR07;CR16;CH17;CO12
Experienced developer	Hard debugging/refactoring	Repro case; approved data route	Trial Astra/Codex local	Fable 5.1/Claude Code	No clear leader; Fable retention/migration limits	medium	Relevant system evidence; incomplete conditions	Repo trials favor alternative	CR01;CR02;CR03;CR04;CR05;CH17;CO04;CO05
Cost-sensitive developer	Bounded volume	Tests; review; approved route/protocol	Pilot GLM-5.3/OpenCode	DeepSeek V4-Pro; GLM Flash	Aliases reroute; Coding Plan privacy unresolved; weight license separate	low	Rates supported; plan policy unresolved	Retries/review erase savings	CGL01;CGL03;CGL04;CGL06;CGL20;CR12;CR17;CO06;CGLP01;CGLP02;CGLP03
Product developer	Greenfield/UI	Browser checks; acceptance criteria	Pilot Gemini 3.8/Antigravity	Existing Sol/Opus workflow	Managed-agent billing unverified; vendor demonstrations	low	Comparative outcomes missing	Measured review and defect costs improve	CR08;CR09
Offline developer	Bounded local tasks	Compatible quant/server; isolated tools; tests/review	Pilot Qwen3.8-27B/4-bit	Smaller Qwen3.5-9B	24GB fit/context/latency untested	low	Weight arithmetic only	Hardware tests fail thresholds	CGL09;CGL17;CGL18;CGL21
Existing GitHub team	Controlled rollout	Approve retention; cap spend	Trial Copilot Business/Astra	Approved Codex Business route	Service retention unknown; staged access	medium	Documented controls; quality unmeasured	Policy or task trial fails	CH02;CH04;CH07;CH12;CH17;CH20;CH21
Engineering lead	Unattended production	Owner; sandbox; rollback	Insufficient evidence for automatic approval	Bounded background tasks; human approval	PR approval preview grants authority	low	No reliable production success rate	Audited trials satisfy acceptance criteria	CH03;CH20;CO07;CO08

Pricing #

provider	plan or model	access route	region	billing basis	monthly usd	input usd per mtok	output usd per mtok	cache read usd per mtok	included usage	notes	claim ids
OpenAI	gpt-6-astra	Direct API	Default route; region-specific terms separate	usage	n/a	10	50	1	No subscription allowance assumed	>272K:2x input/cache, 1.5x output; Fast2x; Batch/Flex50%.	CR01;CR15
Anthropic	claude-fable-5-1	Direct API	Default route; region-specific terms separate	usage	n/a	10	50	0.25	No subscription allowance assumed	5m cache-write shown; 1h20. US-only input/output 1.1x; Batch input/output 50%.	CR03;CR05;CR15
Anthropic	claude-opus-5	Direct API	Default route; region-specific terms separate	usage	n/a	5	25	0.5	No subscription allowance assumed	5m cache-write shown; 1h10.	CR16
Anthropic	claude-sonnet-5	Direct API	Default route; region-specific terms separate	usage	n/a	2	10	0.2	No subscription allowance assumed	5m cache-write shown; 1h4.	CR14
OpenAI	gpt-5.6-sol	Direct API	Default route; region-specific terms separate	usage	n/a	4	20	0.4	No subscription allowance assumed	>272K:2x input/1.5x output; 5 cache-write calculated from 1.25x.	CR13;CR15
Google	gemini-3.8-flash	Direct API	Default route; region-specific terms separate	usage	n/a	0.75	3.75	0.075	No subscription allowance assumed	Storage0.50USD/MTok/hour; output includes thinking; grounding extra.	CR09
DeepSeek	deepseek-v4-pro peak	Direct API	Default route; region-specific terms separate	usage	n/a	1.32	3.96	0.044	Prepaid balance; granted balance used first	Peak Mon–Fri01:00–04:00 and06:00–10:00 UTC; all other hours offpeak.	CR12
DeepSeek	deepseek-v4-flash peak	Direct API	Default route; region-specific terms separate	usage	n/a	0.44	1.32	0.014	Prepaid balance; granted balance used first	Peak Mon–Fri01:00–04:00 and06:00–10:00 UTC; all other hours offpeak.	CR12
Z.ai	GLM-5.3	Z.ai API	unknown	usage	n/a	1.4	4.4	0.26	Usage billed by tokens	Standard posted rates per 1M tokens; taxes/region not established	CGL04
Z.ai	GLM-5.3-Flash standard list	Z.ai API	unknown	usage	n/a	0.15	0.5	0.03	Usage billed by tokens	Standard posted rates per 1M tokens; taxes/region not established	CGL05
Z.ai	Coding Plan Lite advertised starting price	GLM Coding Plan	unknown	other	18	n/a	n/a	n/a	2,000 credits per 5 hours; 10,000 weekly; both caps apply	Advertised starting price is USD 18/month; commitment and account-specific discounts are unknown. Subscription credits do not buy unrestricted general API usage.	CGL06;CGL07
GitHub	Copilot Business	Direct subscription	USD public pricing; regional taxes unknown	monthly	19	n/a	n/a	n/a	1,900 credits/seat/month; billing-entity pool; 1 credit=$0.01	June–August promotion ended September 1; no cheaper-model fallback.	CH07;CH21
GitHub	Copilot Enterprise	Direct subscription	USD public pricing; regional taxes unknown	monthly	39	n/a	n/a	n/a	3,900 credits/seat/month; billing-entity pool; 1 credit=$0.01	GitHub Enterprise Cloud; promotion ended September 1.	CH07;CH21
Cursor	Pro	Direct subscription	USD public pricing; regional taxes unknown	monthly	20	n/a	n/a	n/a	Two model pools; exact quantities unknown	Before tax; do not reuse older dollar allowances.	CH14
Cursor	Pro Plus	Direct subscription	USD public pricing; regional taxes unknown	monthly	60	n/a	n/a	n/a	Two model pools; exact quantities unknown	Before tax; no task-volume equivalence.	CH14
Cursor	Ultra	Direct subscription	USD public pricing; regional taxes unknown	monthly	200	n/a	n/a	n/a	Two model pools; exact quantities unknown	Before tax; no task-volume equivalence.	CH14
Cursor	Teams Standard	Direct subscription	USD public pricing; regional taxes unknown	monthly	40	n/a	n/a	n/a	Two per-seat pools; no user pooling; quantities unknown	Enterprise US residency separately enrolled: 10% eligible-model uplift.	CH11;CH13
Cursor	Teams Premium	Direct subscription	USD public pricing; regional taxes unknown	monthly	120	n/a	n/a	n/a	5x Standard usage; exact quantities unknown	5x allowance is not 5x accepted tasks.	CH13
OpenAI	ChatGPT Plus with Codex	Direct subscription	USD public pricing; regional taxes unknown	monthly	20	n/a	n/a	n/a	Variable; Astra estimated 5–45 local messages/5h	Estimates not fixed limits; context/tools change consumption.	CH17
OpenAI	ChatGPT Business with Codex	Direct subscription	USD public pricing; regional taxes unknown	monthly	25	n/a	n/a	n/a	Variable; standard Astra estimate 5–45 local messages/5h	Business no-training default; retention unknown here.	CH17
OpenAI	ChatGPT Business with Codex annual	Direct subscription	USD public pricing; regional taxes unknown	annual-effective-monthly	20	n/a	n/a	n/a	Variable; standard Business allowance	Annual billing; minimum 2 users; monthly equivalent.	CH17

Privacy and training #

provider	product tier	model	hosting route	region	trains by default	retention	caveats	claim ids
Google	Gemini Developer API Paid Services	gemini-3.8-flash	Google API billing-linked project	Available regions; no residency inference	no	Limited-period abuse logs; duration unspecified; Search grounding30days	Unpaid outside EEA/CH/UK can train; feedback separate	CR10
DeepSeek	Open Platform API downstream route	V4-Pro / V4-Flash	DeepSeek API	Exact API residency not established	unknown	API-wide duration unknown	No blanket API no-training or ZDR commitment established	CR17
OpenAI	Standard API	API models; endpoint exceptions	Direct OpenAI API	unknown	no	Abuse logs up to 30 days with legal/harm exceptions; Responses state 30 days by default	Approved ZDR/MAM differs; store=false is not a blanket no-retention switch.	CH08
OpenAI	ChatGPT Business	Codex-accessible models including Astra	Business workspace	unknown	no	unknown	Business no-training verified; consumer policy and exact service-retention duration not established here.	CH17;CH19
Anthropic	Claude Code commercial Team/Enterprise/API	Claude; additional model exceptions apply	Direct Anthropic commercial account	unknown	no	Standard 30 days; separately approved Enterprise ZDR; local transcripts normally 30 days	Fable-specific retention overrides generic expectations; feedback transcripts five years.	CH05;CH06
Anthropic	Claude consumer Free/Pro/Max	Claude; model exceptions apply	Direct consumer account	unknown	choice	Training on: five years; off: 30 days	Account privacy preference applies to Claude Code; feedback retention separate.	CH05;CH06
Cursor	Privacy Mode; standard ZDR model	Eligible models excluding Fable exceptions	Cursor and selected provider	unknown	no	Provider ZDR with abuse exceptions; temporary encrypted file cache	Cloud Agents and designated non-ZDR models need separate assessment.	CH09;CH10
Cursor	Privacy Mode with Fable opt-in	Claude Fable 5.1	Cursor to Anthropic	unknown	no	Provider retains prompts/outputs; exact duration unknown here	Team-wide approval required; Privacy Mode does not mean Fable ZDR.	CH10
GitHub	Copilot Business/Enterprise	GPT-6 Astra	GitHub; OpenAI/GitHub Azure	unknown	no	OpenAI provider ZDR agreement; GitHub service retention unknown	Provider agreement is not an end-to-end storage guarantee.	CH04;CH12
GitHub	Copilot Business/Enterprise Fable enabled	Claude Fable 5.1	GitHub to Claude hosting providers	unknown	no	Anthropic retention by default; eligible enterprise temporary ZDR through 2026	Off by default; exemptions require approval/internal-operations restrictions.	CH04;CH12;CH18
Z.ai	Enterprise/developer API Services under Additional Terms and DPA	GLM-5.3 / GLM-5.3-Flash; no model-specific exception identified	Z.ai hosted API; applicable enterprise/developer agreement required	Generally Singapore under API DPA	no	DPA says submitted/generated content is not stored on servers; other Customer Data retained for service or legal purposes, with no fixed duration established	Contractual commitments for the defined API scope, not an operational audit. Individual Coding Plan coverage unresolved. Human-access controls and metadata retention duration remain unknown.	CGLP01;CGLP02
Z.ai	Individual GLM Coding Plan	GLM-5.3 / GLM-5.3-Flash	Z.ai hosted Coding Plan endpoint	unknown	unknown	unknown for this route	API terms provide content-use and storage protections for their defined scope. Their application to individual Coding Plan traffic is unverified; obtain account and endpoint-specific confirmation before using restricted repositories.	CGLP01;CGLP02;CGLP03

Client compatibility #

client	model family	execution surface	access route	support type	session portability	notes	claim ids
Codex	DeepSeekV4	CLI; desktop; VSCode extension per vendor	DeepSeek API key	compatible-endpoint	Login-method history groups separate; switching back restores visibility	Vendor documented; not exercised. API ignores background/state and many built-in tools; developer role becomes user.	CR11;CR18
OpenCode	GLM-5.3 / GLM-5.3-Flash	Client dependent; OS not verified	GLM Coding Plan	compatible-endpoint	unknown	Documented support, not observed. Model API protocol restrictions for prior subscribers require an account-specific check.	CGL06;CGL20
Cline	GLM-5.3 / GLM-5.3-Flash	Client dependent; OS not verified	GLM Coding Plan	compatible-endpoint	unknown	Documented support, not observed. Model API protocol restrictions for prior subscribers require an account-specific check.	CGL06;CGL20
Claude Code	GLM-5.3 / GLM-5.3-Flash	Client dependent; OS not verified	GLM Coding Plan	compatible-endpoint	unknown	Documented support, not observed. Model API protocol restrictions for prior subscribers require an account-specific check.	CGL06;CGL20
GitHub Copilot in VS Code	GPT-6 Astra; Claude Fable 5.1	VS Code IDE	Copilot Pro+; Max; Business; Enterprise	native	Continue external Copilot or Claude sessions	Documented August VS Code 1.132–1.135; new-model rollout gradual; Fable policy off by default.	CH01;CH02;CH18
GitHub Copilot CLI/app	GPT-6 Astra; Claude Fable 5.1	CLI; app	Copilot Pro+; Max; Business; Enterprise	native	unknown	Business/Enterprise exclusions GA September 2; assess other tool paths separately.	CH02;CH04;CH18
Cursor	Claude Fable 5.1	Editor; optional Cloud Agents	Cursor account; Fable retention-policy opt-in	native	unknown	Safeguard-triggered Fable requests route to Opus; cloud repositories are stored during execution.	CH10;CH13
Codex	GPT-6 Astra	CLI; IDE (local)	ChatGPT Plus or Business	native	unknown	Astra evidenced in local-message usage table; cloud model eligibility is separate. Check actual sandbox/network policy.	CH17;CH19;CH20
Codex	Models available to API key	CLI; SDK; IDE	OpenAI API key	byo-key	unknown	API-key route excludes cloud GitHub review and Slack; endpoint retention differs.	CH08;CH17;CH19
Claude Code	Claude family; model-specific eligibility	Local CLI; web VM	Claude consumer/commercial account or configured cloud provider	native	Local plaintext transcripts support resumption	First-party and third-party data paths differ; model retention restrictions remain additional.	CH05;CH06
OpenCode	Configured hosted or local model	Local agent; configured provider	BYO key or local endpoint	compatible-endpoint	unknown	Documented only; choose Responses versus Chat Completions adapter; validate tools and context.	CH15
Zed External Agents	Claude; Codex; OpenCode; Copilot; Cursor via ACP	Zed agent panel; separate agent process	Agent-owned provider authentication	native	Cross-agent transfer unknown	No Zed external-agent fee; provider billing and retention separate. Documented, not tested.	CH16

On-prem tiers #

memory type	vram tier	system ram gb	model	quant	practical context	agentic grade	latency	notes	claim ids
dedicated-vram	8 GB capacity-screening candidate	unknown	Qwen3.5-4B	Proposed 4-bit; format unselected	unknown; must measure on selected stack	unknown	unknown; no measurements	Weights-only arithmetic is approximately 2.5 GB using rounded 5B parameters; runtime adds caches, buffers, vision components and scales. No tested fit.	CGL17;CGL18;CGL21
dedicated-vram	16 GB capacity-screening candidate	unknown	Qwen3.5-9B	Proposed 8-bit; format unselected	unknown; must measure on selected stack	unknown	unknown; no measurements	Weights-only arithmetic is approximately 10 GB using rounded 10B parameters; actual runtime and concurrent requests remain unknown.	CGL17;CGL18;CGL21
dedicated-vram	24 GB capacity-screening candidate	unknown	Qwen3.8-27B	Proposed 4-bit; format unselected	unknown; must measure on selected stack	unknown	unknown; no measurements	Weights-only arithmetic is approximately 14 GB. The vendor model score is not this quantization’s score. Native 262K context is not validated usable context here.	CGL09;CGL18;CGL21
unified-memory	64 GB shared-memory capacity-screening candidate	64	Qwen3.8-27B	Proposed 8-bit; format unselected	unknown; must measure on selected stack	unknown	unknown; no measurements	Weights-only arithmetic is approximately 28 GB; the 64 GB shared pool also serves the OS and applications. Engine compatibility and throughput are unverified.	CGL09;CGL18;CGL21
dedicated-vram	Multi-accelerator/server scale	unknown	GLM-5.3	Illustrative 4-bit; released card includes FP8	unknown; must measure on selected stack	unknown	unknown; no measurements	4-bit weights-only arithmetic is approximately 376.5 GB; the released FP8 weights exceed this. No claim that an 8/16/24 GB GPU can fully hold the model.	CGL02;CGL03;CGL18
dedicated-vram	Multi-accelerator/server scale	unknown	GLM-5.3-Flash	Released FP8 mixed types	unknown; must measure on selected stack	unknown	unknown; no measurements	8-bit arithmetic is approximately 321 GB for weights only; 18B activated parameters do not imply 18 GB of total memory.	CGL08;CGL18