Up to date
openai Apache 2.0 out 15 months ago
20.9B parameters (4 of 32 experts per token), text in, text out, 128K tokens of context. 13 providers serve it from $0.018 per million input tokens.
Verified 2026-10-25 Rank #4 in LLMs · score 136.9 method v1 Sources 5 Downloads 30d 5.8M · #25 of 619 tracked
Changed this week: 3 new quantized builds on the Hub (254 in all).2026-10-24 · 1 new quantized build on the Hub (252 in all).2026-10-24 · 1 new provider: Nscale.2026-10-23
Why did the open-weights model cross the road?
to get to the other side of the loss function
__|__
--o--o--(_)--o--o--
/ \
/ \
/ \
/ \
/ \
~~~~~~~~~~~~~
~~~~~~~~~~~~~
~~~~~~~~~~~~~
~~~~~~~~~~~~~
~~~~~~~~~~~~~
~~~~~~~~~~~~~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.3 of 4 pass · parasail-gpt-oss-20b on parasail · run 2026-10-09 · temperature 0 · cost $0.006 · 2026-10 thing: supermarket · all models →
Picked by rule from 47 tracked pieces: the lab's own post, the most shared, a field report, the latest.
OpenAI released gpt-oss-120b and gpt-oss-20b, their first open-weight models since GPT-2 in 2019.
burrito-core is a drop-in inference harness for gpt-oss that provides OpenAI and Anthropic API compatibility.
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.
Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.
10 providers with a live price, from $0.018 to $0.075 per million input tokens; the median is $0.030. One dot per provider, cheapest in teal.
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 (13 API providers); heavily run locally (254 quantized builds); a big base for builders (1,054 fine-tunes, adapters and merges).
Top Spaces running it: pliny-the-prompter/obliteratus ♥ 621 · umint/ai ♥ 258 · deepset/autoquizzer ♥ 226 · bishmoy/Arxiv-CS-RAG ♥ 195
Weights only, an estimate: leave room for the KV cache. Method
About 13 GB at INT4, 25 GB at FP8, 50 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 probe replies as received on 2026-10-07 from parasail, graded by code. Receipts, not a score.
def is_balanced(s: str) -> bool:
stack = []
pairs = {')': '(', ']': '[', '}': '{'}
opens = set(pairs.values())
for ch in s:
if ch in opens:
stack.append(ch)
elif ch in pairs:
if not stack or stack[-1] != pairs[ch]:
return False
amber-falcon-7731
{"city": "Lisbon", "unit": "c"} {"name":"Foo-9B-Instruct","license":"Apache-2.0","parameters_b":9.2,"context_tokens":131072,"languages":["English","French","German","Japanese"]} {
"providers": [
{
"provider": "parasail",
"endpoint_model": "parasail-gpt-oss-20b",
"source_url": "https://api.parasail.io/v1/chat/completions",
"ttft_ms_median": 7726,
"ttft_any_ms_median": 1943,
"tokens_per_s_median": 77.7,
"runs": [
{
"ttft_s": 8.3,
"ttft_any_s": 2.128,
"tokens_per_s": 77.7,
"latency_s": 10.35,
"tokens_out": 804,
"reasoning_chars": 2762,
"finish": "stop"
},
{
"ttft_s": 2.121,
"ttft_any_s": 1.525,
"tokens_per_s": 58.8,
"latency_s": 4.23,
"tokens_out": 249,
"reasoning_chars": 343,
"finish": "stop"
},
{
"ttft_s": 7.726,
"ttft_any_s": 1.943,
"tokens_per_s": 86.6,
"latency_s": 9.26,
"tokens_out": 802,
"reasoning_chars": 2773,
"finish": "stop"
}
]
}
],
"note": "Single client, wall clock; tok/s counts every output token including hidden reasoning."
} Weights, every dated article, and repo commits on one rail.
Teal is the weights (2025-08-04), amber triangles are 7 dated articles (each links to it), grey ticks are repo commits.
1,308 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Everything else, folded: the full tables, every source.
Released under an Apache 2.0 license, this model uses the harmony response format for reasoning and agentic tasks.[1][2] The architecture utilizes 4 experts per token and can run within 16GB of memory using MXFP4 quantization.[3][4]
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.
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.
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | GptOssForCausalLM |
| Model type | gpt_oss |
| Params by dtype | BF16: 1,804,459,584 · U8: 19,110,297,600 |
| Size band | L |
| MoE | yes |
| Experts | 32 |
| Experts per token | 4 |
| Context length | 131,072 (config.json max_position_embeddings) |
| Papers | arXiv:2508.10925 |
| License | Apache 2.0 (apache-2.0) |
| Hub snapshot | 2026-10-25 |
vLLM: supported · GptOssForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · GptOssForCausalLM · 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.
Weights-only memory for gpt-oss-20b: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 50 GB · Single high-end GPU (<80 GB) |
|---|---|
| FP8 (1 byte/param) | 25 GB · Single high-end GPU (<80 GB) |
| INT4 (0.5 bytes/param) | 13 GB · Single GPU (<24 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.
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.
| Probe | Result | Latency | Reply (truncated) |
|---|---|---|---|
| Self-description | info | 6.4 s | I am ChatGPT, a language model based on OpenAI’s GPT‑4 architecture. My training data cutoff is September 2021. |
| Tool call | pass | 2.4 s | {"city": "Lisbon", "unit": "c"} |
| Strict JSON | pass | 4.1 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 | 6.2 s | def is_balanced(s: str) -> bool: stack = [] pairs = {')': '(', ']': '[', '}': '{'} opens = set(pairs.values()) for ch in s: if ch in opens: stack.append(ch) elif ch in pairs: if not stack or stack[-1] != pairs[ch]: return False |
| Multi-step arithmetic | pass | 3.9 s | 16 |
| Needle recall at 128k tokens | pass | 9.4 s | amber-falcon-7731 |
Probe set firstlook.v1 · observed 2026-10-07 20:03 UTC · endpoint · provider docs · cost of this run $0.00
quantized: 254merge: 7finetune: 567adapter: 480
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| mlx-community/gpt-oss-20b-MXFP4-Q8 | quantized | MXFP4 | 20.9B | L | apache-2.0 | mlx-community | 98 | 7.8M |
| OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HF | merge | — | 21.2B | L | apache-2.0 | OpenGVLab | 10 | 6.2M |
| unsloth/gpt-oss-20b-GGUF | quantized | GGUF | — | — | apache-2.0 | unsloth | 891 | 5.3M |
| unsloth/gpt-oss-20b-BF16 | finetune | -BF16 | 20.9B | L | apache-2.0 | unsloth | 35 | 2.8M |
| lmstudio-community/gpt-oss-20b-GGUF | quantized | GGUF | — | — | apache-2.0 | lmstudio-community | 89 | 2.5M |
| unsloth/gpt-oss-20b-unsloth-bnb-4bit | quantized | -bnb | 20.9B | L | apache-2.0 | unsloth | 41 | 2.1M |
| lmstudio-community/gpt-oss-20b-MLX-8bit | quantized | MLX | 20.9B | L | apache-2.0 | lmstudio-community | 60 | 1.8M |
| ggml-org/gpt-oss-20b-GGUF | quantized | GGUF | — | — | apache-2.0 | ggml-org | 184 | 1.6M |
| FastFlowLM/GPT-OSS-20B-NPU2 | quantized | — | — | — | apache-2.0 | FastFlowLM | 1 | 810.2k |
| OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview | merge | — | 392M | S | apache-2.0 | OpenGVLab | 86 | 766.8k |
| unsloth/gpt-oss-20b | quantized | — | 20.9B | L | apache-2.0 | unsloth | 50 | 644k |
| bartowski/openai_gpt-oss-20b-GGUF | quantized | GGUF | — | — | — | bartowski | 39 | 461.8k |
| RedHatAI/gpt-oss-20b | quantized | — | 21.5B | L | apache-2.0 | RedHatAI | 6 | 207k |
| unsloth/gpt-oss-20b-bnb-4bit | quantized | -bnb | 20.9B | L | apache-2.0 | unsloth | 14 | 200.2k |
| ArliAI/gpt-oss-20b-Derestricted | finetune | — | 20.9B | L | apache-2.0 | ArliAI | 98 | 152.8k |
| MuXodious/gpt-oss-20b-RichardErkhov-heresy | finetune | — | 20.9B | L | apache-2.0 | MuXodious | 27 | 116.8k |
| farbodtavakkoli/OTel-LLM-20B-IT | finetune | — | — | — | apache-2.0 | farbodtavakkoli | 1 | 105.2k |
| farbodtavakkoli/OTel-LLM-20B-Reasoning | finetune | — | — | — | apache-2.0 | farbodtavakkoli | 14 | 95.5k |
| Jinx-org/Jinx-gpt-oss-20b | finetune | — | 20.9B | L | apache-2.0 | Jinx-org | 96 | 58.6k |
| amd/gpt-oss-20b-MoE-Quant-W-MXFP4-A-FP8-KV-FP8 | quantized | -Quant | 11.4B | L | apache-2.0 | amd | 3 | 59.8k |
| amd/gpt-oss-20b-WFP8-AFP8-KVFP8 | quantized | FP8- | 20.9B | L | apache-2.0 | amd | 0 | 59.9k |
| DavidAU/Openai_gpt-oss-20b-NEO-GGUF | quantized | GGUF | — | — | apache-2.0 | DavidAU | 25 | 48k |
| squ11z1/gpt-oss-nano | quantized | GGUF | 8.97B | M | apache-2.0 | squ11z1 | 26 | 42.3k |
| mlx-community/gpt-oss-20b-MXFP4-Q4 | quantized | MXFP4 | 20.9B | L | apache-2.0 | mlx-community | 19 | 40.1k |
| Mungert/DMind-3-GGUF | quantized | GGUF | — | — | apache-2.0 | Mungert | 0 | 37.2k |
| Mungert/gpt-oss-20b-Derestricted-GGUF | quantized | GGUF | — | — | apache-2.0 | Mungert | 3 | 35.6k |
| chromadb/context-1 | finetune | — | 20.9B | L | apache-2.0 | chromadb | 423 | 10.9k |
| wangzhang/gpt-oss-20b-abliterated | finetune | — | 20.9B | L | apache-2.0 | wangzhang | 5 | 31.7k |
| bartowski/openai_gpt-oss-20b-GGUF-MXFP4-Experimental | quantized | GGUF | — | — | — | bartowski | 9 | 29.5k |
| mlx-community/gpt-oss-20b-OptiQ-4bit | quantized | -4bit | 20.9B | L | apache-2.0 | mlx-community | 14 | 28.2k |
| DavidAU/Openai_gpt-oss-20b-CODER-NEO-CODE-DI-MATRIX-GGUF | quantized | GGUF | — | — | apache-2.0 | DavidAU | 20 | 25.1k |
| DevQuasar/openai.gpt-oss-20b-GGUF | quantized | GGUF | — | — | — | DevQuasar | 2 | 20.6k |
| InferenceIllusionist/gpt-oss-20b-MLX-4bit | quantized | MLX | 20.9B | L | apache-2.0 | InferenceIllusionist | 20 | 19.3k |
| 2imi9/gpt-oss-20B-NVFP4A16-BF16 | quantized | -BF16 | 20.9B | L | apache-2.0 | 2imi9 | 5 | 15.7k |
| KJML/gpt-oss-20b-FP8-Dynamic | quantized | -FP8 | 20.9B | L | apache-2.0 | KJML | 1 | 15.3k |
| EricSpencer00/chattla-20b | quantized | GGUF | 20.9B | L | apache-2.0 | EricSpencer00 | 1 | 15.1k |
| manjunathshiva/gpt-oss-20b-tq3 | quantized | — | 2.94B | M | apache-2.0 | manjunathshiva | 47 | 12.4k |
| giladgd/gpt-oss-20b-GGUF | quantized | GGUF | — | — | apache-2.0 | giladgd | 2 | 14.6k |
| m-polignano/ANITA-NEXT-20B-gpt-oss-ITA | finetune | — | 20.9B | L | apache-2.0 | m-polignano | 5 | 14.4k |
| fesalfayed/gpt-oss-20b-hermes_agent-tool-finetune_gguf | quantized | gguf | — | — | apache-2.0 | fesalfayed | 6 | 14.3k |
| Date | Commit | SHA |
|---|---|---|
| 2025-08-26 | Update README.md | 6cee5e8 |
| 2025-08-13 | Configure `<|call|>` as an EOS token (#109) | d666cf3 |
| 2025-08-13 | Update README.md | f4770b2 |
| 2025-08-08 | Update chat_template.jinja (#94) | cbf31f6 |
| 2025-08-08 | Update README.md | db6852c |
| 2025-08-07 | Updated chat template (#82) | 10e9d71 |
| 2025-08-06 | Update the model name in USAGE_POLICY (#52) | 2e8f805 |
| 2025-08-06 | Remove chat_template.json (#59) | ef533c9 |
| 2025-08-05 | Update README | bbf0930 |
| 2025-08-05 | Update LICENSE | f47b956 |
| 2025-08-05 | Update vLLM command | 6cd4d0f |
| 2025-08-05 | Add model card | 21ae1e7 |
| 2025-08-05 | Upload folder using huggingface_hub | ec854da |
| 2025-08-05 | Upload folder using huggingface_hub | 5a5e34e |
| 2025-08-04 | Upload folder using huggingface_hub | f81fef1 |
| 2025-08-04 | initial commit | 2923e0b |
Hosted endpoints observed for gpt-oss-20b, 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 |
|---|---|---|---|---|---|---|---|---|---|---|
| Darkbloom openrouter | openai/gpt-oss-20b | 128K | fp8 | $0.02 | $0.09 | — | — | yes | live | 2026-10-24 |
| AkashML openrouter | openai/gpt-oss-20b | 128K | fp4 | $0.02 | $0.10 | — | — | yes | live | 2026-10-24 |
| DeepInfra hf, deepinfra, openrouter | openai/gpt-oss-20b | 128K | bfloat16 | $0.02 | $0.11 | 127.9 | 307 ms | yes | live | 2026-10-25 |
| DekaLLM openrouter | openai/gpt-oss-20b | 128K | bf16 | $0.03 | $0.14 | — | — | yes | live | 2026-10-24 |
| Parasail parasail, openrouter | parasail-gpt-oss-20b | 128K | fp4 | $0.03 | $0.15 | — | — | yes | serverless | 2026-10-24 |
| CoreWeave openrouter | openai/gpt-oss-20b | 128K | fp4 | $0.03 | $0.13 | — | — | yes | live | 2026-10-24 |
| SiliconFlow openrouter | openai/gpt-oss-20b | 128K | fp8 | $0.04 | $0.18 | — | — | — | status -2 | 2026-10-24 |
| Nscale hf | openai/gpt-oss-20b | 128K | — | $0.05 | $0.20 | 171.7 | 311 ms | yes | live | 2026-10-25 |
| OVHcloud hf | gpt-oss-20b | 128K | — | $0.05 | $0.18 | 36.6 | 441 ms | yes | live | 2026-10-25 |
| Amazon Bedrock openrouter | openai/gpt-oss-20b | 128K | unknown | $0.07 | $0.15 | — | — | yes | live | 2026-10-24 |
| Google openrouter | openai/gpt-oss-20b | 128K | unknown | $0.07 | $0.25 | — | — | — | status -5 | 2026-10-24 |
| Groq hf, openrouter | openai/gpt-oss-20b | 128K | unknown | $0.07 | $0.30 | 721 | 425 ms | yes | live | 2026-10-25 |
| Featherless hf | openai/gpt-oss-20b | — | — | — | — | — | — | — | live | 2026-10-25 |
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.
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
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
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.