Step-3.7-Flash

201B parameters (8 of 288 experts per token), images and text in, text out, 256K tokens of context. 4 providers serve it from $0.20 per million input tokens.

Verified 2026-10-25 Rank #324 in LLMs · score 82.2 method v1 Sources 4 Downloads 30d 16k · #475 of 619 tracked

Can I run it?121GB at INT4Estimatefits 2×80 GBmethod

What does it cost?$0.20/1M incheapest of 3DeepInfra2026-10-25

Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24

Do people use it?16kdownloads, 30 days#475 of 619 we track · below the median+1 likes, downloads −19% this week463 likesHugging Face2026-10-24

Worth reading

Picked by rule from 11 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Step 3.7 Flash achieves 92.82% F1 on DeepSearchQA, comparable to Kimi K2.6, a 1T / 32B-active model.

From Step 3.7 Flash · static.stepfun.com · 2026-05-29

The lab's own post

Step 3.7 Flash

Step 3.7 Flash is a high-efficiency agentic model featuring native multimodal understanding and enhanced web search capabilities.

Step 3.7 Flash achieves 92.82% F1 on DeepSearchQA, comparable to Kimi K2.6, a 1T / 32B-active model.
The model reaches 45.8% on GDPval across 44 occupations and passes over 98% on Tau2-bench Telecom.

static.stepfun.com 2026-05-29 48 HN pts · 16 comments 2.2k words sells inference

Latest · 5 months ago

Step 3.7 Flash loves to use wait in the reasoning

Step 3.7 Flash loves to use "wait" in the reasoning.

gist.github.com 2026-06-02 2 HN pts 867 words

11 best of 11 tracked

The 3 picks above, plus these 8. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleDeepSeek V4 Flash 0731 (Max) vs Step 3.7 Flashartificialanalysis.ai · undated
Article阶跃星辰stepfun.com · undated · official
Codestepfun-ai/Step-3.7-Flashgithub.com · undated
ArticleStep-3.7-Flash - AI Model | HowAIWorks.aihowaiworks.ai · 2026-05-29
ArticleStepFun Step 3.7 Flash: 198B Model at 400 tok/s for $0.20/M Input (2026)aimadetools.com · 2026-05-29
ArticleStep 3.7 Flash Review: We Tested StepFun's 198B Model ...flowtivity.ai · undated
ArticleStepFun Releases Step 3.7 Flash: A 198B MoE Vision- ...marktechpost.com · undated
ArticleStep 3.7 Flash: The New Agent Model Explained | Logic & Layerslogicandlayers.com · undated

Show all 8

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

3 providers from $0.20

3 providers with a live price, from $0.20 to $0.20 per million input tokens; the median is $0.20. One dot per provider, cheapest in teal.

#	Cheapest 4	In /1M	Out /1M	Context	Quant	Seen
1	DeepInfra	$0.20	$1.15	256K	—	Oct 25
2	Novita	$0.20	$1.15	256K	fp8	Oct 24
3	StepFun	$0.20	$1.15	250K	fp8	Oct 24
4	Featherless	—	—	—	—	Oct 25

All 4 providers →

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

Served by 4 API providers; 41 quantized builds for local use.

As a service4DeepInfra, Featherless, Novita, StepFun
On own hardware41quantized builds
Built on1010 fine-tunes · 0 adapters · 0 merges
In apps26Hugging Face Spaces

Top Spaces running it: stepfun-ai/Step-3.7-Flash-dev ♥ 17 · akhaliq/Step-3.7-Flash ♥ 6 · ranilmukesh/PlacementPredictor ♥ 1 · RehamAAhmed/Ai_Inspection_System ♥ 1

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits two 80 GB GPUs at INT4; too big for one

About 121 GB at INT4, 242 GB at FP8, 483 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2026-05-23), amber triangles are 5 dated articles (each links to it), grey ticks are repo commits.

What people built on it

51 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

51 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF unsloth/Step-3.7-Flash-GGUF for llama.cpp, Ollama, LM Studio 714.1k downloads MLX mlx-community/Step-3.7-Flash-4bit for Apple silicon 4.5k downloads AWQ cyankiwi/Step-3.7-Flash-AWQ-INT4 for vLLM, SGLang 18.1k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

8 rows

Benchmarks 8 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (8)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported ClawEval-1.1	67.1	—	model author	2026-10-07
author-reported GDPVal-AA	45.8	—	model author	2026-10-07
author-reported HLE w. Tool	48.1	—	model author	2026-10-07
author-reported SWE-bench Pro · Write code with agents	56.3	—	model author	2026-10-07
author-reported SimpleVQA (Search)	79.2	—	model author	2026-10-07
author-reported Terminal-Bench · Write code with agents	59.5	—	model author	2026-10-07
author-reported Toolathlon · Call tools and APIs	49.5	—	model author	2026-10-07
author-reported V* (Python)	95.3	—	model author	2026-10-07

What the card says

2 cited sentences

Summary generated, cited

Step 3.7 Flash is a 198B-parameter sparse Mixture-of-Experts vision-language model.^[1] The architecture combines a 196B-parameter language backbone with a 1.8B-parameter vision encoder for native image understanding.^[2]

“198B-parameter sparse Mixture-of-Experts (MoE) vision-lan…”
“combines a 196B-parameter language backbone with a 1.8B-p…”

Generated 2026-10-07 15:06 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

6 use cases

What it is good at 6 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

23 fields

Listed in LLMs · MoE

Facts

Step-3.7-Flash has 201B parameters. (safetensors metadata, observed 2026-10-24)
Step-3.7-Flash is a mixture-of-experts model with 288 experts, routing 8 per token. (config.json, observed 2026-10-24)
Context length is 262,144 tokens (256K). (config.json max_position_embeddings, observed 2026-10-24)
License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
Created 2026-05-23, last modified 2026-06-03. (HF API, observed 2026-10-24)

Pipeline tag	image-text-to-text
Library	transformers
Architectures	Step3p7ForConditionalGeneration
Model type	step3p7
Dtype	bfloat16
Params by dtype	BF16: 201,365,304,064 · F32: 12,096
Size band	XXL
MoE	yes
Experts	288
Experts per token	8
Context length	262,144 (config.json max_position_embeddings)
Languages	en
License	Apache 2.0 (apache-2.0)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Step3p7ForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Step3p7ForConditionalGeneration · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Step-3.7-Flash: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	483 GB · Multi-GPU node (<640 GB)
FP8 (1 byte/param)	242 GB · Multi-GPU node (<640 GB)
INT4 (0.5 bytes/param)	121 GB · Multi-GPU node (<640 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

40 variants

Lineage

Original model Family: stepfun-ai/Step-3.7 51 derivatives on the Hub

quantized: 41finetune: 10

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
unsloth/Step-3.7-Flash-GGUF	quantized	GGUF	—	—	apache-2.0	unsloth	66	714.1k
stepfun-ai/Step-3.7-Flash-GGUF	quantized	GGUF	—	—	apache-2.0	stepfun-ai	176	141.9k
bartowski/Step-3.7-Flash-GGUF	quantized	GGUF	—	—	apache-2.0	bartowski	8	45.3k
tarruda/Step-3.7-Flash-GGUF	quantized	GGUF	—	—	apache-2.0	tarruda	13	20.6k
cyankiwi/Step-3.7-Flash-AWQ-INT4	quantized	AWQ	43.4B	XL	apache-2.0	cyankiwi	6	18.1k
mudler/Step-3.7-Flash-APEX-GGUF	quantized	GGUF	—	—	apache-2.0	mudler	15	15.2k
AesSedai/Step-3.7-Flash-GGUF	quantized	GGUF	—	—	—	AesSedai	31	7.9k
notSnix/Step-3.7-Flash-Q4_K_M-MTP-GGUF	quantized	GGUF	—	—	apache-2.0	notSnix	14	5.8k
mlx-community/Step-3.7-Flash-4bit	quantized	-4bit	197B	XXL	apache-2.0	mlx-community	5	4.5k
notSnix/Step-3.7-Flash-MTP-Draft-GGUF	quantized	GGUF	—	—	apache-2.0	notSnix	6	4.4k
lemuralabs/Step-3.7-Flash-OptiQ-3.7bpw-mlx	quantized	7bpw	30.0B	L	apache-2.0	lemuralabs	3	2.6k
a4lg/Step-3.7-Flash-MTP-ONLY-GGUF	quantized	GGUF	—	—	apache-2.0	a4lg	1	2.1k
Hikari07jp/Step-3.7-Flash-MTP-draft	finetune	—	2.96B	M	apache-2.0	Hikari07jp	6	1.7k
notSnix/Step-3.7-Flash-Q4_K_M-GGUF	quantized	GGUF	—	—	apache-2.0	notSnix	1	1.9k
mlx-community/Step-3.7-Flash-8bit	quantized	-8bit	197B	XXL	apache-2.0	mlx-community	1	1.5k
DevQuasar/stepfun-ai.Step-3.7-Flash-GGUF	quantized	GGUF	—	—	—	DevQuasar	0	1.4k
AEON-7/Step-3.7-Flash-AEON-Ultimate-Abliterated-BF16	finetune	-BF16	201B	XXL	apache-2.0	AEON-7	3	1.1k
lemuralabs/Step-3.7-Flash-MXFP4-mlx	quantized	MXFP4	201B	XXL	apache-2.0	lemuralabs	0	1.2k
olka-fi/Step-3.7-Flash-MXFP4	quantized	MXFP4	112B	XXL	apache-2.0	olka-fi	2	946
BIGJUTT/Step-3.7-Flash-GGUF	quantized	GGUF	—	—	apache-2.0	BIGJUTT	1	957
Baekpica/Step-3.7-Flash-Mixed-Quant-GGUF	quantized	-Quant	—	—	apache-2.0	Baekpica	2	826
ashuaria/Step-3.7-Flash-GGUF-FineTune	quantized	GGUF	—	—	apache-2.0	ashuaria	0	859
ibrahimkettaneh/Step-3.7-Flash-uncensored-abliterated-heretic-BF16	finetune	-BF16	201B	XXL	other	ibrahimkettaneh	7	438
mlx-community/Step-3.7-Flash-6bit	quantized	—	197B	XXL	apache-2.0	mlx-community	0	783
AigustOfwind/Step-3.7-Flash-JANG_2L-CRACK	quantized	—	310B	XXL	other	AigustOfwind	0	751
AxionML/Step-3.7-Flash-NVFP4	quantized	NVFP4	104B	XXL	apache-2.0	AxionML	0	748
mlx-community/Step-3.7-Flash-oQ3e	quantized	—	199B	XXL	apache-2.0	mlx-community	1	631
ivanfioravanti/Step-3.7-Flash-3bit-gs32	quantized	—	199B	XXL	apache-2.0	ivanfioravanti	1	524
mlx-community/Step-3.7-Flash-oQ2e	quantized	—	199B	XXL	apache-2.0	mlx-community	0	535
dawncr0w/Step-3.7-Flash-oQ3.5-MLX	quantized	MLX	199B	XXL	apache-2.0	dawncr0w	0	524
ivanfioravanti/Step-3.7-Flash-4bit	quantized	-4bit	199B	XXL	apache-2.0	ivanfioravanti	0	505
kingjones777/Step-3.7-Flash-ROCmFP4-STRIX_LEAN-GGUF	quantized	GGUF	—	—	apache-2.0	kingjones777	2	388
tcclaviger/Step-3.7-Flash-240REAP-MXFP416	quantized	—	99.8B	XL	apache-2.0	tcclaviger	2	366
dawncr0w/Step-3.7-Flash-oQ3-MLX	quantized	MLX	199B	XXL	apache-2.0	dawncr0w	0	453
dawncr0w/Step-3.7-Flash-oQ4-MLX	quantized	MLX	199B	XXL	apache-2.0	dawncr0w	0	449
anerjy/Step-3.7-Flash-MLX-adapter	finetune	MLX	—	—	apache-2.0	anerjy	0	437
dawncr0w/Step-3.7-Flash-oQ8-MLX	quantized	MLX	199B	XXL	apache-2.0	dawncr0w	0	401
dawncr0w/Step-3.7-Flash-oQ2-MLX	quantized	MLX	199B	XXL	apache-2.0	dawncr0w	0	374
turboderp/Step-3.7-Flash-exl3	quantized	exl3	—	—	apache-2.0	turboderp	5	108
tcclaviger/Step-3.7-Flash-240REAP	finetune	—	170B	XXL	apache-2.0	tcclaviger	1	285

Repository changelog

23 commits

Repository changelog 23 commits

Date	Commit	SHA
2026-06-03	add cache_position to mask_kwargs in modeling_step3p7.py (#13)	5f62440
2026-06-01	delete unused parameters `use_qk_norm`	f711925
2026-05-29	Update README.md	751ea53
2026-05-29	Fix repo name casing in SGLang FP8 example	5371d3c
2026-05-29	Update README.md	f69ad14
2026-05-29	update Ecosystem bullet	63f1ea4
2026-05-28	Update README.md	5c5f97c
2026-05-28	update benchmark chart	a214531
2026-05-28	fix vLLM intro to match StepFun-specific Docker tag	f0bbde3
2026-05-28	update Local Deployment install instructions	e7debd8
2026-05-28	sync to v3 model card	e03d01d
2026-05-28	sync to v2 model card	fbcee77
2026-05-28	describe MoE as sparse for precision	8817d4a
2026-05-28	clarify regional base_url and use env vars in examples	ee8c807
2026-05-28	add benchmark chart above Pricing section	1678751
2026-05-28	clean up model card: fix numbering, typos, and code examples	47ec689
2026-05-28	add readme (#1)	480473d
2026-05-26	update processor config	a9c0171
2026-05-23	add step-3.7-flash bf16 model libs	7805a18
2026-05-23	add step-3.7-flash bf16 model libs	c171c6a
2026-05-23	add step-3.7-flash bf16 model config	457483d
2026-05-23	step-3.7-flash bf16 model	6578499
2026-05-23	initial commit	dc3047b

All providers

4 providers · 4 endpoints

Where to run 4 providers · 4 endpoints

Hosted endpoints observed for Step-3.7-Flash, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.20/M.

Provider	Model ref	Context	Quant	$/M in	$/M out	tok/s	TTFT	Tools	Status	Observed
DeepInfra hf	stepfun-ai/Step-3.7-Flash	256K	—	$0.20	$1.15	17.5	1358 ms	yes	live	2026-10-25
Novita openrouter	stepfun/step-3.7-flash	256K	fp8	$0.20	$1.15	—	—	yes	live	2026-10-24
StepFun openrouter	stepfun/step-3.7-flash	250K	fp8	$0.20	$1.15	—	—	yes	live	2026-10-24
Featherless hf	stepfun-ai/Step-3.7-Flash	—	—	—	—	—	—	—	live	2026-10-25

Articles and coverage

11 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official2

Step 3.7 Flash

static.stepfun.com2026-05-29officialHN 48 pts · 16 comments2,216 words

Step 3.7 Flash is a high-efficiency agentic model featuring native multimodal understanding and enhanced web search capabilities. Compared to Step 3.5 Flash, it gains +5% on SWE-Bench Pro and 6.1% on Terminal-Bench 2.1. The model reaches 97% of Claude Opus 4.6's coding performance at roughly one-ninth the per-task cost via Advisor Mode, which consults a larger model only at inflection points. It uses a Python tool for cropping, zooming, and drawing to enable high-resolution visual reasoning and GUI operation.written, reviewed, cited

Step 3.7 Flash achieves 92.82% F1 on DeepSearchQA, comparable to Kimi K2.6, a 1T / 32B-active model.
The model reaches 45.8% on GDPval across 44 occupations and passes over 98% on Tau2-bench Telecom.
Step 3.7 Flash achieves 75.82% on BrowseComp, approaching larger models such as Claude Opus 4.7 and GLM 5.1.

阶跃星辰

stepfun.comofficial5,044 words

Step 3.7 Flash is a high-efficiency model for real-world agents capable of up to 400 TPS with native multimodal understanding and reliable tool orchestration. It achieves 59.5 on Terminal-Bench 2.1 and supports Advisor Mode, which consults a larger advisor model only at inflection points like planning or recovery. This mode reaches 97% of Claude Opus 4.6's coding performance at roughly one-ninth the per-task cost ($0.19 v.s. $1.76 per task).written, reviewed, cited

Step 3.7 Flash achieves 59.5 on Terminal-Bench 2.1, outperforming DeepSeek V4 Flash's 53.4 on the same metric.
The model reaches 95.3 on V*(with Python), trailing only Kimi K2.6's 96.9 among listed multimodal benchmarks.
Step 3.7 Flash scores 45.8 on GDPval across 44 occupations, while passing over 98% of Tau2-bench Telecom reasoning tiers.

Analysis8

Step 3.7 Flash loves to use wait in the reasoning

gist.github.com2026-06-02authorityHN 2 pts · 0 comments867 words

Step 3.7 Flash loves to use "wait" in the reasoning. The model concludes that driving is better than walking for a car wash 50 meters away because walking would require returning to the car to bring it to the wash, adding unnecessary extra movement. The 50m drive is short and delivers the car directly to the wash with no extra steps.written, reviewed, cited

DeepSeek V4 Flash 0731 (Max) vs Step 3.7 Flash

artificialanalysis.aiauthority844 words

DeepSeek V4 Flash 0731 (Max) scores 34 on the Artificial Analysis Intelligence Index compared to Step 3.7 Flash's estimated 19. DeepSeek V4 Flash 0731 (Max) costs $0.0721 per 1M tokens versus $0.183 for Step 3.7 Flash and features a 1000k token context window against 262k tokens. Step 3.7 Flash generates 192 tokens per second with a 2.37s time to first token, while DeepSeek V4 Flash 0731 (Max) outputs 150 tokens per second with a 1.29s time to first token.written, reviewed, cited

stepfun-ai/Step-3.7-Flash

github.comauthority1,736 words

Step 3.7 Flash is a 198B-parameter sparse MoE vision-language model combining a 196B-parameter language backbone with a 1.8B-parameter vision encoder for native image understanding. It activates approximately 11B parameters per token, delivering throughput up to 400 tokens per second while supporting a 256k context window and three selectable reasoning levels. The model is available on OpenRouter and NVIDIA NIM, with input pricing at $0.20 / M tokens for cache misses and $1.15 / M tokens for output, while local deployment requires at least 128GB unified memory.written, reviewed, cited

The NVFP4 quantized variant requires modelopt activation and FP8 KV Cache alignment for efficient deployment.
Running the NVFP4 checkpoint on a GB200 TP=4 setup improves aggregate decode throughput by up to 1.45x at 64 concurrency.
Deployment of this model requires transformers 5.0 or later to function correctly.

StepFun Step 3.7 Flash Review (2026): We Tested the 1/9-Cost-of-Opus Claim | DataLLM Lab

datallmlab.com2026-06-18

Step-3.7-Flash - AI Model | HowAIWorks.ai

howaiworks.ai2026-05-292,441 words

StepFun Step 3.7 Flash: 198B Model at 400 tok/s for $0.20/M Input (2026)

aimadetools.com2026-05-291,349 words

Step 3.7 Flash Review: We Tested StepFun's 198B Model ...

flowtivity.ai2,681 words

StepFun Releases Step 3.7 Flash: A 198B MoE Vision- ...

marktechpost.com

All coverage tracked for this model (11) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
DeepSeek V4 Flash 0731 (Max) vs Step 3.7 Flash	artificialanalysis.ai	—	2026-10-07	exa
阶跃星辰	stepfun.com official	—	2026-10-07	exa
stepfun-ai/Step-3.7-Flash	github.com	—	2026-10-07	exa
Step 3.7 Flash Review: We Tested StepFun's 198B Model ...	flowtivity.ai	—	2026-10-07	exa
StepFun Releases Step 3.7 Flash: A 198B MoE Vision- ...	marktechpost.com	—	2026-10-07	exa
Step 3.7 Flash: The New Agent Model Explained \| Logic & Layers	logicandlayers.com	—	2026-10-07	exa
StepFun Step 3.7 Flash Review (2026): We Tested the 1/9-Cost-of-Opus Claim \| DataLLM Lab	datallmlab.com	2026-06-18	2026-10-07	exa
Step 3.7 Flash loves to use wait in the reasoning	gist.github.com	2026-06-02	2026-10-07	hn
Step 3.7 Flash	static.stepfun.com official	2026-05-29	2026-10-07	hn, exa
Step-3.7-Flash - AI Model \| HowAIWorks.ai	howaiworks.ai	2026-05-29	2026-10-07	exa
StepFun Step 3.7 Flash: 198B Model at 400 tok/s for $0.20/M Input (2026)	aimadetools.com	2026-05-29	2026-10-07	exa

Similar models

3 models

Similar models

Qwen3.8-Flash-Next · 180B
same size band (XXL), also MoE, same task (image-text-to-text), fewer providers
Qwen3-VL-235B-A22B-Instruct · 236B
same size band (XXL), also MoE, same task (image-text-to-text), permissive license
Qwen3-VL-235B-A22B-Thinking · 236B
same size band (XXL), also MoE, same task (image-text-to-text), permissive license

Sources

4 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:37 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:06 UTC · Last verified 2026-10-25

Follow Step-3.7-Flash

Watch this model

/feeds/models/stepfun-ai--Step-3.7-Flash.xml(RSS 2.0) · every change to Step-3.7-Flash
/feeds/families/stepfun-ai--Step-3.7.xml(RSS 2.0) · every change across the Step-3.7 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.