Qwen3.8-27B

27.8B parameters, images and text in, text out, 256K tokens of context. 23 providers serve it from $0.035 per million input tokens.

Verified 2026-10-25 Rank #1 in LLMs · score 140.2 method v1 Sources 5 Downloads 30d 7.9M · #18 of 619 tracked

Changed this week: 99 new quantized builds on the Hub (1,542 in all).2026-10-24 · Cheapest input price fell 42.9%: $0.035 → $0.020 per 1M (OpenRouter).2026-10-24 · 98 new quantized builds on the Hub (1,541 in all).2026-10-24

Can I run it?17GB at INT4Estimatefits a 24 GB cardmethodCommunity: Run QWEN3.8 27B on 16gb Nvidia GPUs →

What does it cost?$0.035/1M incheapest of 21Darkbloomfp42026-10-24

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

Does it work?5/5probes · 87 tok/sfirst lookparasail2026-10-07

Do people use it?7.9Mdownloads, 30 days#18 of 619 we track · top 3%+434 likes, downloads +1% this week18,314 likesHugging Face2026-10-24

SVG reply: a pelican steals a bicycle from a supermarket — Draw a pelican stealing a bicycle from a supermarket. Reply with one complete SVG document only, no prose and no code fences. Use viewBox="0 0 200 150" and a transparent background. Use only these colours for every fill and stroke: #06e0c0 #ffc298 #cd85f9 #f1f2f2.

Finish the dad joke

as received● 1 line · 10 words

Why did the open-weights model cross the road?

To get to the other side and show its weights

Finish this dad joke. Reply with the punchline only, on one line, without quotes: Why did the open-weights model cross the road?

A parasailer, in ASCII

as received● 12×19 fits 12×32

       .-""""-.
      /        \
     /          \
      \        /
       \      /
        \    /
         \  /
          \/
          o
         /|\
        / | \
   ~~~~~  /_  ~~~~~

Draw ASCII art of a parasailer over the sea. At most 12 lines and 32 columns. Reply with the art only: no code fences, no explanation.

Animate a bouncing ball

as received● 8 frames · 16×16

Make a 16x16 pixel sprite animation of a ball bouncing, 8 frames. Reply with JSON only, no code fences: {"palette": {"a": "#rrggbb", ...}, "fps": <number>, "frames": [[16 strings of 16 characters], ...]}. At most 4 palette colours with single-letter keys; "." means transparent; exactly 8 frames.

The ringer: the same four playful tests for every model, replies shown as received and checked by code. A joke, not a benchmark; never ranked.4 of 4 pass · parasail-qwen38-27b on parasail · run 2026-10-09 · temperature 0 · cost $0.074 · 2026-10 thing: supermarket · all models →

Worth reading

Picked by rule from 42 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Qwen3.8-27B-Humanlike-Chat 2.0: texts like a human, now with tool calls and better instruction following

reddit.com 2026-10-02

Latest · 38 days ago

Run QWEN3.8 27B on 16gb Nvidia GPUs

github.com 2026-09-17 28 HN pts · 10 comments 6.2k words

25 best of 42 tracked

The 3 picks above, plus these 22. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleQwen3.8-Flash-Nextqwen.ai · 2026-08-26 · official
ArticleQwen3.8-Max: A New Bar for Coding and Coworkqwen.ai · 2026-08-03 · official
ArticleQwen3.8 27B scores 52 on Artificial Analysisartificialanalysis.ai · 2026-08-17 · 381 HN pts
ArticleQwen3.8 Max now ranked as the best overall model by agentic indexartificialanalysis.ai · 2026-08-06
CodeShow HN: Running 104GB Qwen3.8-Flash-Next on 48GB Mac with at ~12 tok/sgithub.com · 2026-09-01
ArticleBenchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapsesquesma.com · 2026-08-26 · 286 HN pts
ArticleQwen3.8-Flash-Next Intelligence, Performance and Price Analysisartificialanalysis.ai · 2026-08-27
DiscussionQwen3.8-27Btwitter.com · 2026-08-14 · 299 HN pts
ArticleQwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiencyqwen.ai · 2026-08-26 · official
ArticleQwen3.8-Maxqwencloud.com · 2026-09-02
ArticleQwen3.8 27B at 256K: 50 TPS on a 24 GB GPUpiszczek.pl · 2026-08-17 · 38 HN pts
CodeQwen3.8-Flash-Next Technical Report [pdf]github.com · 2026-08-26
CodeQwen3.8-27B make medium the default effort level instead of xhighgithub.com · 2026-08-19 · 16 HN pts
ArticleQwen3.8 Weights Releasedmodelscope.cn · 2026-08-12
DiscussionAsk HN: Which harness do you use with Qwen3.8 27B? And why?news.ycombinator.com · 2026-09-23 · 4 HN pts
CodeI found why Qwen3.8-27B was 14× slower on my 4070 Supergithub.com · 2026-08-23 · 8 HN pts
ArticleI Rented a 96 GB GPU and Took Uncensored Qwen3.8 From 44 to 125 tok/saseemshrey.com · 2026-08-24
ArticleTuning a Local Coding Agent: Oh My Pi and Qwen3.8-27B on Two RTX 3090sdoug.sh · 2026-09-15
DiscussionQwen 3.8-Max and Qwen3.8-27B are going open weighttwitter.com · 2026-08-03
CodePi vs. the official DeepSeek Harness on the same local model (Qwen3.8-27B)github.com · 2026-08-25
ArticleThinkingCap-Qwen3.8-27B: Less Thinking, Similar Accuracykaitchup.substack.com · 2026-09-26
DiscussionQwen3.8 Fetches Weird URLsnews.ycombinator.com · 2026-08-20

Show all 22

Where to run it

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

Prices for the same weights vary 28×

21 providers with a live price, from $0.035 to $0.99 per million input tokens; the median is $0.20. One dot per provider, cheapest in teal.

#	Cheapest 5	In /1M	Out /1M	Context	Quant	Seen
1	Darkbloom	$0.035	$2.20	256K	fp4	Oct 24
2	Chutes	$0.039	$1.25	256K	fp8	Oct 24
3	Near AI	$0.040	$1.35	256K	fp8	Oct 24
4	Ionstream	$0.040	$1.95	256K	fp8	Oct 24
5	Reka	$0.040	$2.20	256K	fp8	Oct 24

All 23 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.

Widely served (23 API providers); heavily run locally (1,542 quantized builds); a big base for builders (1,000 fine-tunes, adapters and merges).

As a service23AkashML, Alibaba, Cerebras, Chutes +19
On own hardware1,542quantized builds
Built on1,000721 fine-tunes · 254 adapters · 25 merges
In apps100+Hugging Face Spaces

Top Spaces running it: RioShiina/ImageGen ♥ 222 · JonathanColetti/Qwen3.8-27B-Uncensored-Demo ♥ 209 · embedl/hfviewer ♥ 181 · victor/Qwen3.8-27B-free-endpoint ♥ 101

Will it fit?

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

Fits a 64 GB Mac (unified memory) at INT4

About 17 GB at INT4, 33 GB at FP8, 67 GB at BF16: weights only, so leave room for the KV cache.

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

Day-one receipts

The probe replies as received on 2026-10-07 from parasail, graded by code. Receipts, not a score.

Write a function● pass

def is_balanced(s: str) -> bool:
    stack = []
    matching = {')': '(', ']': '[', '}': '{'}
    for ch in s:
        if ch in '([{':
            stack.append(ch)
        elif ch in matching:
            if not stack or stack.pop() != matching[ch]:
                return False
    return not stack

Needle at 128K● pass

amber-falcon-7731

Tool call● pass

{"city": "Lisbon", "unit": "c"}

Strict JSON● pass

{
  "name": "Foo-9B-Instruct",
  "license": "Apache-2.0",
  "parameters_b": 9.2,
  "context_tokens": 131072,
  "languages": [
    "English",
    "French",
    "German",
    "Japanese"
  ]
}

Speed run● pass

{
  "providers": [
    {
      "provider": "parasail",
      "endpoint_model": "parasail-qwen38-27b",
      "source_url": "https://api.parasail.io/v1/chat/completions",
      "ttft_ms_median": 11694,
      "ttft_any_ms_median": 578,
      "tokens_per_s_median": 93.3,
      "runs": [
        {
          "ttft_s": null,
          "ttft_any_s": 0.71,
          "tokens_per_s": 110.3,
          "latency_s": 13.6,
          "tokens_out": 1500,
          "reasoning_chars": 4408,
          "finish": "length"
        },
        {
          "ttft_s": null,
          "ttft_any_s": 0.578,
          "tokens_per_s": 93.3,
          "latency_s": 16.08,
          "tokens_out": 1500,
          "reasoning_chars": 4672,
          "finish": "length"
        },
        {
          "ttft_s": 11.694,
          "ttft_any_s": 0.495,
          "tokens_per_s": 75,
          "latency_s": 13.31,
          "tokens_out": 998,
          "reasoning_chars": 2692,
          "finish": "stop"
        }
      ]
    }
  ],
  "note": "Single client, wall clock; tok/s counts every output token including hidden reasoning."
}

Three-tool errand● pass

{
  "calls": [
    {
      "name": "get_weather",
      "args": {
        "city": "Lisbon",
        "date": "2026-10-14"
      }
    },
    {
      "name": "check_calendar",
      "args": {
        "date": "2026-10-14"
      }
    },
    {
      "name": "book_table",
      "args": {
        "restaurant": "Cervejaria Ramiro",
        "date": "2026-10-14",
        "time": "19:30",
        "party_size": 2
      }
    }
  ],
  "checks": {
    "get_weather": true,
    "check_calendar": true,
    "book_table": true
  },
  "final": "Done! Here's the summary:\n\n- **Weather in Lisbon on Wed 2026-10-14:** Clear, 5% chance of rain, high of 22°C — no rain expected.\n- **Free evening slot:** 19:30–21:30\n- **Booking:** Table for 2 at Cervejaria Ramiro at 19:30\n\n**Confirmation code: TBL-4471** 🍺",
  "fixture": "receipts.errand.v1"
}

Every probe, with timings and cost →

The story so far

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

10 dated articles and 16 repo commits so far

Teal is the weights (2026-08-05), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

What people built on it

2,542 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

2,542 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/Qwen3.8-27B-GGUF for llama.cpp, Ollama, LM Studio 20.7M downloads MLX lmstudio-community/Qwen3.8-27B-MLX-4bit for Apple silicon 10.2M downloads FP8 Qwen/Qwen3.8-27B-FP8 for vLLM, SGLang 14.3M downloads

All variants →

Reference

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

Benchmarks

5 rows

Benchmarks 5 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (5)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 DeepSWE · Write code with agents	42.2	—	model author	2026-10-07
author-reported NL2Repo · Write code with agents	42.3	—	model author	2026-10-07
author-reported SWE-bench · Write code with agents	79	—	model author	2026-10-07
author-reported SWE-bench Pro · Write code with agents	61.7	—	model author	2026-10-07
author-reported Terminal-Bench · Write code with agents	73	—	model author	2026-10-07

What the card says

1 cited sentence

Summary generated, cited

This dense model functions as a native vision-language model for understanding images and videos.^[1]

“a native vision-language model that understands images an…”

Generated 2026-10-07 14:59 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

20 fields

Listed in LLMs

Facts

Qwen3.8-27B has 27.8B parameters. (safetensors metadata, 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 2026-08-05, last modified 2026-08-14. (HF API, observed 2026-10-24)

Pipeline tag	image-text-to-text
Library	transformers
Architectures	Qwen3_5ForConditionalGeneration
Model type	qwen3_5
Dtype	bfloat16
Params by dtype	BF16: 27,781,427,952
Size band	L
MoE	no
Context length	262,144 (config.json max_position_embeddings)
License	Apache 2.0 (apache-2.0)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Qwen3_5ForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Qwen3_5ForConditionalGeneration · 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 Qwen3.8-27B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	67 GB · Single high-end GPU (<80 GB)
FP8 (1 byte/param)	33 GB · Single high-end GPU (<80 GB)
INT4 (0.5 bytes/param)	17 GB · Single GPU (<24 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

First look, every probe

6 probes

First look measured

A fixed probe set run against a live API endpoint: the same questions for every model, answers graded by code, nothing judged by a model. Latencies are single-request wall clock from one client location, not a benchmark.

via parasail parasail-qwen38-27b 5/5 probes pass first visible token 28.6 s (thinking starts at 0.7 s) end-to-end 87 tok/s (median of 3)

Probe	Result	Latency	Reply (truncated)
Self-description	info	2.9 s	I am Qwen, a large language model developed by Alibaba Group’s Tongyi Lab. My training data cutoff is 2026.
Tool call	pass	4.8 s	{"city": "Lisbon", "unit": "c"}
Strict JSON	pass	3.9 s	{ "name": "Foo-9B-Instruct", "license": "Apache-2.0", "parameters_b": 9.2, "context_tokens": 131072, "languages": [ "English", "French", "German", "Japanese" ] }
Code (hidden tests)	pass	3.3 s	def is_balanced(s: str) -> bool: stack = [] matching = {')': '(', ']': '[', '}': '{'} for ch in s: if ch in '([{': stack.append(ch) elif ch in matching: if not stack or stack.pop() != matching[ch]: return False return not stack
Multi-step arithmetic	pass	1.6 s	16
Needle recall at 128k tokens	pass	10.1 s	amber-falcon-7731

Probe set firstlook.v1 · observed 2026-10-07 20:18 UTC · endpoint · provider docs · cost of this run $0.05

All variants

40 variants

Lineage

Original model Family: Qwen/Qwen3.8 2,542 derivatives on the Hub

quantized: 1,542adapter: 254finetune: 721merge: 25

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
unsloth/Qwen3.8-27B-GGUF	quantized	GGUF	—	—	apache-2.0	unsloth	5.3k	20.7M
Qwen/Qwen3.8-27B-FP8	quantized	-FP8	27.8B	L	apache-2.0	Qwen	941	14.3M
lmstudio-community/Qwen3.8-27B-MLX-4bit	quantized	MLX	27.4B	L	apache-2.0	lmstudio-community	112	10.2M
lmstudio-community/Qwen3.8-27B-MLX-8bit	quantized	MLX	27.4B	L	apache-2.0	lmstudio-community	50	9.8M
lmstudio-community/Qwen3.8-27B-MLX-6bit	quantized	MLX	27.4B	L	apache-2.0	lmstudio-community	38	9.7M
lmstudio-community/Qwen3.8-27B-MLX-5bit	quantized	MLX	27.4B	L	apache-2.0	lmstudio-community	16	9.6M
cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF	quantized	iMatrix	—	—	apache-2.0	cdiamond	43	7.6M
JonathanColetti/Qwen3.8-27B-Uncensored-GGUF	quantized	GGUF	—	—	apache-2.0	JonathanColetti	1.7k	6.5M
unsloth/Qwen3.8-27B-NVFP4	quantized	NVFP4	19.9B	L	apache-2.0	unsloth	522	6.4M
huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF	quantized	GGUF	—	—	apache-2.0	huihui-ai	1.5k	6.1M
prism-ml/Ternary-Bonsai-2-27B-gguf	quantized	gguf	—	—	apache-2.0	prism-ml	2.8k	5.4M
HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF	quantized	GGUF	—	—	apache-2.0	HauhauCS	2.1k	5.3M
lmstudio-community/Qwen3.8-27B-GGUF	quantized	GGUF	—	—	apache-2.0	lmstudio-community	76	4.7M
ggml-org/Qwen3.8-27B-GGUF	quantized	GGUF	—	—	apache-2.0	ggml-org	137	4.1M
RadixArk/Qwen3.8-27B-NVFP4	quantized	NVFP4	18.2B	L	apache-2.0	RadixArk	118	3.9M
0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF	quantized	GGUF	—	—	apache-2.0	0bserverx	748	3.5M
ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF	quantized	GGUF	—	—	apache-2.0	ISTA-DASLab	2.3k	2.8M
cyankiwi/Qwen3.8-27B-AWQ-INT4	quantized	AWQ	27.8B	L	apache-2.0	cyankiwi	123	2.8M
autotrust/JEV-27B-VL	adapter	—	27.8B	L	apache-2.0	autotrust	1.2k	2.4M
peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF	quantized	GGUF	—	—	apache-2.0	peculiar-ragdoll	246	2.3M
byteshape/Qwen3.8-27B-GGUF	quantized	GGUF	—	—	apache-2.0	byteshape	180	2.2M
Inferact/Qwen3.8-27B-NVFP4	quantized	NVFP4	17.6B	L	apache-2.0	Inferact	21	2.1M
OBLITERATUS/Qwen3.8-27B-OBLITERATED	quantized	GGUF	27.8B	L	apache-2.0	OBLITERATUS	1.5k	2M
gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090	quantized	NVFP4	14.6B	L	apache-2.0	gittensor-model-hub	173	1.3M
z-lab/Qwen3.8-27B-DFlash2-GGUF	quantized	GGUF	—	—	apache-2.0	z-lab	155	1.1M
bartowski/Qwen3.8-27B-GGUF	quantized	GGUF	—	—	apache-2.0	bartowski	214	1.1M
mlx-community/Qwen3.8-27B-4bit	quantized	-4bit	27.4B	L	apache-2.0	mlx-community	224	1.1M
empero-ai/Qwen3.8-27B-Ridge-GGUF	quantized	GGUF	—	—	apache-2.0	empero-ai	363	998.2k
RedHatAI/Qwen3.8-27B-INT4	quantized	INT4	27.4B	L	apache-2.0	RedHatAI	57	862.7k
incoai/Qwen3.8-27B-DFlash2	finetune	—	1.92B	S	apache-2.0	incoai	289	850.4k
agentionai/Qwen3.8-27B-AP-GGUF	quantized	GGUF	—	—	apache-2.0	agentionai	93	808.8k
nvidia/Qwen3.8-27B-NVFP4	quantized	NVFP4	18.2B	L	apache-2.0	nvidia	211	752.7k
sakamakismile/Qwen3.8-27B-MTP-NVFP4	quantized	NVFP4	27.4B	L	apache-2.0	sakamakismile	45	743.8k
philbert440/Qwen3.8-27B-W4A16-AWQ	quantized	-W4A16	27.4B	L	apache-2.0	philbert440	32	739.3k
orcarouter/Qwen3.8-27B-Uncensored-FP8	quantized	-FP8	27.8B	L	apache-2.0	orcarouter	1.8k	642.3k
unsloth/Qwen3.8-27B	finetune	—	27.8B	L	apache-2.0	unsloth	55	597.9k
orcarouter/Qwen3.8-27B-Uncensored-GGUF	quantized	GGUF	—	—	apache-2.0	orcarouter	1.6k	482.1k
neroued/Qwen3.8-27B-NInfer	quantized	—	—	—	apache-2.0	neroued	80	532.7k
neroued/Qwen3.8-27B-nvfp4-NInfer	quantized	nvfp4	—	—	apache-2.0	neroued	113	496.9k
z-lab/Qwen3.8-27B-DFlash2	finetune	—	1.92B	S	apache-2.0	z-lab	318	484.6k

Repository changelog

16 commits

changed in the last 7 daysPrice moves: 2026-10-24Other: 2026-10-24, 2026-10-24all changes

Repository changelog 16 commits

Date	Commit	SHA
2026-08-14	Update README.md	1d4bf0f
2026-08-14	Update README.md	ed1e737
2026-08-14	Update README.md	0044ccc
2026-08-14	Update README.md (model card)	9206f8d
2026-08-14	Update README.md	08cb375
2026-08-14	Update README.md (model card)	99b86b4
2026-08-14	Add LICENSE	e13a4f0
2026-08-13	Upload generation_config.json with huggingface_hub	dbdc473
2026-08-13	Upload chat_template.jinja with huggingface_hub	412f8b6
2026-08-13	Upload tokenizer_config.json with huggingface_hub	452438e
2026-08-13	Upload folder using huggingface_hub	72a217a
2026-08-13	Update README.md	69b8fb7
2026-08-13	Update README.md	3bb536d
2026-08-12	Update README.md	e744676
2026-08-12	Update README.md	4191656
2026-08-05	initial commit	6714f56

All providers

23 providers · 23 endpoints

Where to run 23 providers · 23 endpoints

Hosted endpoints observed for Qwen3.8-27B, 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: $0.02/M.

Provider	Model ref	Context	Quant	$/M in	$/M out	tok/s	TTFT	Tools	Status	Observed
Tenstorrent openrouter	qwen/qwen3.8-27b	256K	fp4	$0.02	$0.50	—	—	yes	status -2	2026-10-24
Darkbloom openrouter	qwen/qwen3.8-27b	256K	fp4	$0.04	$2.20	—	—	yes	live	2026-10-24
Chutes openrouter	qwen/qwen3.8-27b	256K	fp8	$0.04	$1.25	—	—	yes	live	2026-10-24
Near AI openrouter	qwen/qwen3.8-27b	256K	fp8	$0.04	$1.35	—	—	yes	live	2026-10-24
Ionstream openrouter	qwen/qwen3.8-27b	256K	fp8	$0.04	$1.95	—	—	yes	live	2026-10-24
Reka openrouter	qwen/qwen3.8-27b	256K	fp8	$0.04	$2.20	—	—	yes	live	2026-10-24
Modal openrouter	qwen/qwen3.8-27b	256K	unknown	$0.04	$2.30	—	—	yes	live	2026-10-24
DekaLLM openrouter	qwen/qwen3.8-27b	256K	unknown	$0.04	$2.50	—	—	yes	live	2026-10-24
Wafer openrouter	qwen/qwen3.8-27b	256K	fp8	$0.06	$2.30	—	—	yes	live	2026-10-24
DeepInfra hf, deepinfra, openrouter	Qwen/Qwen3.8-27B	256K	bf16	$0.15	$1.88	34.7	419 ms	yes	live	2026-10-25
Mancer 2 openrouter	qwen/qwen3.8-27b	256K	fp8	$0.18	$2.50	—	—	yes	live	2026-10-24
ModelRun openrouter	qwen/qwen3.8-27b	256K	fp4	$0.20	$2.50	—	—	yes	live	2026-10-24
Phala openrouter	qwen/qwen3.8-27b	977K	unknown	$0.20	$2.50	—	—	yes	live	2026-10-24
AkashML openrouter	qwen/qwen3.8-27b	256K	fp8	$0.23	$1.98	—	—	yes	live	2026-10-24
Parasail parasail, openrouter	parasail-qwen38-27b	256K	fp8	$0.24	$2.20	—	—	yes	serverless	2026-10-24
CoreWeave openrouter	qwen/qwen3.8-27b	256K	fp8	$0.40	$3.00	—	—	yes	live	2026-10-24
Novita hf, openrouter	qwen/qwen3.8-27b	977K	unknown	$0.42	$3.00	40.5	968 ms	yes	live	2026-10-25
Alibaba openrouter	qwen/qwen3.8-27b	977K	unknown	$0.42	$2.55	—	—	yes	live	2026-10-24
Cloudflare openrouter	qwen/qwen3.8-27b	256K	unknown	$0.45	$3.20	—	—	yes	live	2026-10-24
Venice openrouter	qwen/qwen3.8-27b	256K	fp8	$0.45	$3.20	—	—	yes	live	2026-10-24
OVHcloud hf	Qwen3.8-27B	256K	—	$0.47	$3.19	39.9	3749 ms	yes	live	2026-10-25
Cerebras hf, openrouter	qwen-3.8-27b	64K	fp16	$0.99	$1.49	994.5	205 ms	yes	live	2026-10-25
Featherless hf	Qwen/Qwen3.8-27B	—	—	—	—	—	—	—	live	2026-10-25

Articles and coverage

42 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.

Analysis8

Qwen3.8 27B scores 52 on Artificial Analysis

artificialanalysis.ai2026-08-17authorityHN front pageHN 381 pts · 180 comments986 words

Qwen3.8 27B (Xhigh) is a multimodal reasoning model from Alibaba released on August 14, 2026, that scores 34 on the Artificial Analysis Intelligence Index. It generated 200M output tokens during evaluation, costing $0.50 per 1M input tokens and $3.00 per 1M output tokens. The model outputs text at 43 tokens per second and is licensed under Apache 2.0.written, reviewed, cited

The model's 256k token context window is explicitly defined as approximately 384 A4 pages of size 12 Arial font.
For blended pricing calculations, the analysis uses a specific 7:2:1 ratio for cache hits, inputs, and outputs.
Reasoning models are compared across both reasoning and non-reasoning categories rather than just within their own class.

Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses

quesma.com2026-08-26HN front pageHN 286 pts · 139 comments1,154 words

Run Qwen3.8 27B locally: real numbers from my Mac Studio

terminalbytes.com2026-08-28HN front pageHN 140 pts · 99 comments2,021 words

Run QWEN3.8 27B on 16gb Nvidia GPUs

github.com2026-09-17authorityHN 28 pts · 10 comments6,217 words

Qwen3.8 27B at 256K: 50 TPS on a 24 GB GPU

piszczek.pl2026-08-17HN 38 pts · 36 comments

Qwen3.8-27B make medium the default effort level instead of xhigh

github.com2026-08-19authorityHN 16 pts · 4 comments

GVS5H: Five Qwen3.8-27B Models Match Claude Fable 5 on LiveCodeBench Hard

github.com2026-09-12authorityHN 6 pts · 0 comments

I found why Qwen3.8-27B was 14× slower on my 4070 Super

github.com2026-08-23authorityHN 8 pts · 0 comments

Community2

Qwen3.8-27B

twitter.com2026-08-14HN front pageHN 299 pts · 3 comments

Ask HN: Which harness do you use with Qwen3.8 27B? And why?

news.ycombinator.com2026-09-23HN 4 pts · 10 comments

All coverage tracked for this model (42) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
Qwen 3.8-27B — The Architecture, From Scratch	atandra2000.github.io	—	2026-10-07	exa
Qwen3.8-27B: The Dense 27B That Carries the Max Lineage	boxwrench.dev	—	2026-10-07	exa
ThinkingCap-Qwen3.8-27B: Less Thinking, Similar Accuracy	kaitchup.substack.com	2026-09-26	2026-10-07	hn
Ask HN: Which harness do you use with Qwen3.8 27B? And why?	news.ycombinator.com	2026-09-23	2026-10-07	hn
Run QWEN3.8 27B on 16gb Nvidia GPUs	github.com	2026-09-17	2026-10-07	hn
Tuning a Local Coding Agent: Oh My Pi and Qwen3.8-27B on Two RTX 3090s	doug.sh	2026-09-15	2026-10-07	hn
GVS5H: Five Qwen3.8-27B Models Match Claude Fable 5 on LiveCodeBench Hard	github.com	2026-09-12	2026-10-07	hn
Qwen3.8-Max	qwencloud.com	2026-09-02	2026-10-07	hn
Show HN: Running 104GB Qwen3.8-Flash-Next on 48GB Mac with at ~12 tok/s	github.com	2026-09-01	2026-10-07	hn
Qwen3.8 27B: specs, benchmarks, and how to run it on Lyceum \| Lyceum	lyceum.technology	2026-09-01	2026-10-07	exa
Agentic SQL for Free with Qwen3.8 27B and DuckDB	motherduck.com	2026-08-29	2026-10-07	hn
Qwen3.8-27B and Radeon R9700 and 13-year-old Pentium = actual good performance	mateusznowak.dev	2026-08-28	2026-10-07	hn
Run Qwen3.8 27B locally: real numbers from my Mac Studio	terminalbytes.com	2026-08-28	2026-10-07	hn
Qwen3.8-Flash-Next Intelligence, Performance and Price Analysis	artificialanalysis.ai	2026-08-27	2026-10-07	hn
Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses	quesma.com	2026-08-26	2026-10-07	hn
Qwen3.8-Flash-Next Technical Report [pdf]	github.com	2026-08-26	2026-10-07	hn
Qwen3.8-Flash-Next	qwen.ai official	2026-08-26	2026-10-07	hn
Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiency	qwen.ai official	2026-08-26	2026-10-07	hn
Pi vs. the official DeepSeek Harness on the same local model (Qwen3.8-27B)	github.com	2026-08-25	2026-10-07	hn
I Rented a 96 GB GPU and Took Uncensored Qwen3.8 From 44 to 125 tok/s	aseemshrey.com	2026-08-24	2026-10-07	hn
I found why Qwen3.8-27B was 14× slower on my 4070 Super	github.com	2026-08-23	2026-10-07	hn
Qwen3.8 Fetches Weird URLs	news.ycombinator.com	2026-08-20	2026-10-07	hn
Qwen3.8-27B make medium the default effort level instead of xhigh	github.com	2026-08-19	2026-10-07	hn
Qwen3.8 27B Model review by Kodesage	kodesage.ai	2026-08-19	2026-10-07	exa
Qwen3.8-27B on a single RTX 3090: crash fix, 131K context, 9 myths	jonidimo.github.io	2026-08-17	2026-10-07	hn
We're hosting Qwen3.8-27B for Free API usage	sprag.ai	2026-08-17	2026-10-07	hn
Qwen3.8 27B scores 52 on Artificial Analysis	artificialanalysis.ai	2026-08-17	2026-10-07	hn
Qwen3.8 27B at 256K: 50 TPS on a 24 GB GPU	piszczek.pl	2026-08-17	2026-10-07	hn
Qwen 3.8 27B Release: A Deep Dive	kie.ai	2026-08-17	2026-10-07	exa
Deep Dive into Qwen3.8-27B Hybrid Attention Architecture: How 48 Gated DeltaNet + 16 Attention Layers Make a 27B Model Run on Consumer GPUs \| HappyRock	happyrock.cloud	2026-08-16	2026-10-07	exa
Qwen3.8-27B – Release Day Demos	loktar00.github.io	2026-08-15	2026-10-07	hn
Qwen3.8 27B Coming to Cerebras	twitter.com	2026-08-15	2026-10-07	hn
Qwen3.8-27B Review: The Open-Weight 27B Worth Running	orcarouter.ai	2026-08-15	2026-10-07	exa
Qwen3.8-27B Explained: Hybrid Attention, 262K Context, New Benchmarks \| MindStudio	mindstudio.ai	2026-08-15	2026-10-07	exa
Qwen3.8-27B	twitter.com	2026-08-14	2026-10-07	hn
Qwen/Qwen3.8-27B	kinonn-bot.github.io	2026-08-14	2026-10-07	exa
Qwen3.8 27B, First Impressions - peaster	peaster.io	2026-08-14	2026-10-07	exa
Qwen3.8 Weights Released	modelscope.cn	2026-08-12	2026-10-07	hn
Qwen3.8 Max now ranked as the best overall model by agentic index	artificialanalysis.ai	2026-08-06	2026-10-07	hn
Qwen 3.8-Max and Qwen3.8-27B are going open weight	twitter.com	2026-08-03	2026-10-07	hn
Qwen3.8-Max: A New Bar for Coding and Cowork	qwen.ai official	2026-08-03	2026-10-07	hn
Qwen3.8-Max	twitter.com	2026-08-03	2026-10-07	hn

Similar models

3 models

Similar models

Qwen3.6-27B · 27.8B
same size band (L), same task (image-text-to-text), fewer providers, permissive license
Qwen3.5-27B · 27.8B
same size band (L), same task (image-text-to-text), fewer providers, permissive license
gemma-3-27b-it · 27.4B
same size band (L), same task (image-text-to-text), fewer providers

Sources

5 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:31 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 14:59 UTC · Last verified 2026-10-25

Follow Qwen3.8-27B

Watch this model

/feeds/models/Qwen--Qwen3.8-27B.xml(RSS 2.0) · every change to Qwen3.8-27B
/feeds/families/Qwen--Qwen3.8.xml(RSS 2.0) · every change across the Qwen3.8 family

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

The ringer: four playful tests

A pelican steals a bicycle from a supermarket

Finish the dad joke

A parasailer, in ASCII

Animate a bouncing ball

Worth reading

Qwen3.8-27B-Humanlike-Chat 2.0: texts like a human, now with tool calls and better instruction following

Run Qwen3.8 27B locally: real numbers from my Mac Studio

GVS5H: Five Qwen3.8-27B Models Match Claude Fable 5 on LiveCodeBench Hard

Run QWEN3.8 27B on 16gb Nvidia GPUs

25 best of 42 tracked

Where to run it

Prices for the same weights vary 28×

Where it's used

Will it fit?

Fits a 64 GB Mac (unified memory) at INT4

Day-one receipts

The story so far

10 dated articles and 16 repo commits so far

What people built on it

2,542 derivatives on the Hub, most of them quantized

Reference

Benchmarks

Benchmarks 5 rows

What the card says

Summary generated, cited

Use-case evidence

What it is good at 6 use cases

Facts

Facts

Servability

Servability

Hardware fit

Hardware fit ESTIMATE

First look, every probe

First look measured

All variants

Lineage

Top variants (40 shown, ranked by downloads)

Repository changelog

Repository changelog 16 commits

All providers

Where to run 23 providers · 23 endpoints

Articles and coverage

Analysis8

Community2

Similar models

Similar models

Sources

Sources

Follow Qwen3.8-27B

Watch this model