daniicruzz commited on
Commit
616168a
·
verified ·
1 Parent(s): 81c6504

Add model card

Browse files
Files changed (1) hide show
  1. README.md +458 -0
README.md ADDED
@@ -0,0 +1,458 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ base_model:
6
+ - MultiverseComputingCAI/Pulsar-16B
7
+ - nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
8
+ tags:
9
+ - fp8
10
+ - quantized
11
+ - modelopt
12
+ - compactifai
13
+ ---
14
+
15
+ # Pulsar 16B — FP8
16
+ <div align="center">
17
+
18
+ [![License](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
19
+ [![HuggingFace](https://img.shields.io/badge/🤗-Model_Hub-yellow.svg)](TODO_PULSAR_HF_URL)
20
+ [![Discord](https://img.shields.io/badge/Discord-Community-5865F2?logo=discord&logoColor=white)](https://discord.gg/cGas9uStqp)
21
+
22
+ Powered by CompactifAI · Quantized with NVIDIA TensorRT Model Optimizer
23
+
24
+ **Optimized for Fast and Efficient Inference** · **Reduced Memory Footprint** · **FP8 Post-Training Quantization**
25
+
26
+ </div>
27
+
28
+ ---
29
+
30
+ ## Table of Contents
31
+
32
+ - [Model Overview](#model-overview)
33
+ - [Key Characteristics](#key-characteristics)
34
+ - [Quantization](#quantization)
35
+ - [Quick Start](#quick-start)
36
+ - [Reasoning Control](#thinking-reasoning-control)
37
+ - [Tool Calling](#tool-calling)
38
+ - [Training & Fine-Tuning](#training--fine-tuning)
39
+ - [Evaluation & Benchmarks](#evaluation--benchmarks)
40
+ - [Languages](#languages)
41
+ - [Safety & Limitations](#safety--limitations)
42
+ - [Model Information](#model-information)
43
+ - [Citation](#citation)
44
+
45
+ ---
46
+
47
+ ## Model Overview
48
+
49
+ **Pulsar 16B FP8** is the **FP8-quantized release** of [Pulsar-16B](https://huggingface.co/MultiverseComputingCAI/Pulsar-16B), a model based on [NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16), developed by **Multiverse Computing**. The original model is a **~31.6B parameter**, part of the Nemotron model family. It supports **long-context inference up to 1M tokens** and is designed for general-purpose language modeling tasks.
50
+
51
+ The BF16 release applies **model compression techniques** to significantly reduce parameter count and deployment requirements while maintaining compatibility with the Nemotron Hybrid Mamba2-Transformer with MoE architecture. The resulting model achieves **50% compression**, reducing the parameter count to **16.15B parameters** and lowering memory requirements. **This checkpoint additionally applies FP8 post-training quantization** on top of the compressed model.
52
+
53
+ ---
54
+
55
+ ## Key Characteristics
56
+
57
+ | Characteristic | Description |
58
+ |-----------------------|-------------|
59
+ | Base model | [nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16). **31.6B** total parameters, **3.6B** activated per forward pass (11.34% activation ratio). [NVIDIA Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). |
60
+ | Pulsar (BF16 release) | [MultiverseComputingCAI/Pulsar-16B](https://huggingface.co/MultiverseComputingCAI/Pulsar-16B). **16.15B** total parameters, **3.1B** activated per forward pass (19.28% activation ratio) after CompactifAI compression. |
61
+ | Pulsar (this model) | **16.15B** parameters quantized to **FP8** (weights & activations). KV cache stored in **FP8**. |
62
+ | 📐 **Architecture** | Hybrid Mamba2-Transformer with MoE (same family as the base checkpoint). |
63
+ | 🛠️ **Tool calling** | Yes. Same tool-call structure and format as [Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16). See [Tool Calling](#tool-calling). |
64
+ | 🗜️ **Compression** | CompactifAI (proprietary compression technology) |
65
+ | 🔢 **Quantization** | FP8 post-training quantization via [NVIDIA TensorRT Model Optimizer](https://github.com/NVIDIA/TensorRT-Model-Optimizer). |
66
+ | Primary language | English |
67
+
68
+ ---
69
+
70
+ ## Quantization
71
+
72
+ This checkpoint is a **post-training-quantized (PTQ)** version of [Pulsar-16B](https://huggingface.co/MultiverseComputingCAI/Pulsar-16B) produced with [NVIDIA TensorRT Model Optimizer](https://github.com/NVIDIA/TensorRT-Model-Optimizer) (`modelopt` v0.42.0).
73
+
74
+ | Field | Value |
75
+ |---------------------|-------|
76
+ | Quantization scheme | **FP8** — weights & activations |
77
+ | KV cache | **FP8** |
78
+ | Producer | `modelopt` 0.42.0 |
79
+ | Excluded modules | `lm_head`, all Mamba2 `conv1d` modules, and the attention/Mamba projections (`q/k/v/o_proj`, `in_proj`, `out_proj`) of selected layers (kept in higher precision). |
80
+
81
+ ---
82
+
83
+ ## Quick Start
84
+ This model can be loaded with the **Transformers** API. Use `trust_remote_code=True`. Recommended approach: `AutoModelForCausalLM` with `apply_chat_template`. This configuration has been tested with Transformers 4.57.6.
85
+
86
+ ```python
87
+ import torch
88
+ from transformers import AutoTokenizer, AutoModelForCausalLM
89
+
90
+ model_id = "MultiverseComputingCAI/Pulsar-16B-FP8"
91
+
92
+ tokenizer = AutoTokenizer.from_pretrained(
93
+ model_id,
94
+ trust_remote_code=True
95
+ )
96
+
97
+ model = AutoModelForCausalLM.from_pretrained(
98
+ model_id,
99
+ device_map="cuda" if torch.cuda.is_available() else "auto",
100
+ trust_remote_code=True,
101
+ )
102
+ messages = [
103
+ {"role": "user", "content": "Write a haiku about GPUs"},
104
+ ]
105
+
106
+ tokenized_chat = tokenizer.apply_chat_template(
107
+ messages,
108
+ tokenize=True,
109
+ add_generation_prompt=True,
110
+ return_tensors="pt"
111
+ ).to(model.device)
112
+
113
+ outputs = model.generate(
114
+ tokenized_chat,
115
+ max_new_tokens=1024,
116
+ temperature=1.0,
117
+ top_p=1.0,
118
+ eos_token_id=tokenizer.eos_token_id
119
+ )
120
+ print(tokenizer.decode(outputs[0]))
121
+ ```
122
+ Alternatively you can use the `pipeline` API with `trust_remote_code=True`; the pipeline returns the full conversation structure, so extract the assistant message from `outputs[0]["generated_text"]` as needed.
123
+
124
+ ### vLLM Serving
125
+
126
+ #### Installation
127
+
128
+ ```bash
129
+ pip install -U "vllm>=0.12.0"
130
+ ```
131
+
132
+ #### Reasoning parser (NVIDIA)
133
+
134
+ Pulsar uses the same Nemotron v3 reasoning tags as the base model. NVIDIA provides the vLLM plugin as [`nano_v3_reasoning_parser.py`](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/blob/main/nano_v3_reasoning_parser.py) on the base Hugging Face repo (not specific to Pulsar). Direct download:
135
+
136
+ ```bash
137
+ wget https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/main/nano_v3_reasoning_parser.py
138
+ ```
139
+
140
+ You can keep any local filename; the `vllm serve` flags below assume the file is in the current directory as `nano_v3_reasoning_parser.py`. If you mirror an identical copy under the Pulsar model repo, use that URL instead.
141
+
142
+ #### Serve
143
+
144
+ ```bash
145
+ vllm serve MultiverseComputingCAI/Pulsar-16B-FP8 \
146
+ --served-model-name model \
147
+ --max-num-seqs 8 \
148
+ --tensor-parallel-size 1 \
149
+ --port 8000 \
150
+ --trust-remote-code \
151
+ --enable-auto-tool-choice \
152
+ --tool-call-parser qwen3_coder \
153
+ --reasoning-parser-plugin nano_v3_reasoning_parser.py \
154
+ --reasoning-parser nano_v3
155
+ ```
156
+
157
+ > **Note:** The NeMo container `nvcr.io/nvidia/nemo:25.11.nemotron_3_nano` comes with `mamba_ssm` and `causal-conv1d` pre-installed.
158
+
159
+ ---
160
+
161
+ ## Thinking (Reasoning) Control
162
+
163
+ Pulsar supports a **hybrid reasoning mode**: the model can either think step-by-step before answering (reasoning mode) or reply directly (non-reasoning mode). The behaviour is controlled via the `enable_thinking` flag in the chat template.
164
+
165
+ > This section provides a brief overview of reasoning control in Pulsar. For comprehensive details please see the official Nemotron-3 Nano-30B model card at: https://build.nvidia.com/nvidia/nemotron-3-nano-30b-a3b/modelcard
166
+
167
+
168
+ ---
169
+
170
+ ### Transformers API
171
+
172
+ Pass `enable_thinking` through `apply_chat_template`:
173
+
174
+ **Thinking ON (default)**
175
+ ```python
176
+ tokenized_chat = tokenizer.apply_chat_template(
177
+ messages,
178
+ tokenize=True,
179
+ add_generation_prompt=True,
180
+ return_tensors="pt",
181
+ enable_thinking=True, # default — can be omitted
182
+ )
183
+ ```
184
+
185
+ **Thinking OFF**
186
+ ```python
187
+ tokenized_chat = tokenizer.apply_chat_template(
188
+ messages,
189
+ tokenize=True,
190
+ add_generation_prompt=True,
191
+ return_tensors="pt",
192
+ enable_thinking=False,
193
+ )
194
+ ```
195
+
196
+ When thinking is ON the model opens a `<think>` block before the answer.
197
+
198
+ ```python
199
+ output = tokenizer.decode(outputs[0], skip_special_tokens=True)
200
+ # Split on </think> to separate reasoning from the final answer
201
+ if "</think>" in output:
202
+ reasoning, answer = output.split("</think>", 1)
203
+ reasoning = reasoning.replace("<think>", "").strip()
204
+ answer = answer.strip()
205
+ else:
206
+ answer = output
207
+ ```
208
+
209
+ ---
210
+
211
+ ### vLLM
212
+
213
+ #### Server-level default
214
+
215
+ Set the default for **all requests** at startup with `--default-chat-template-kwargs`.
216
+
217
+ > Requires recent versions of vLLM.
218
+
219
+ **Thinking OFF for all requests**
220
+ ```bash
221
+ vllm serve MultiverseComputingCAI/Pulsar-16B-FP8 \
222
+ --served-model-name model \
223
+ --reasoning-parser-plugin nano_v3_reasoning_parser.py \
224
+ --reasoning-parser nano_v3 \
225
+ --trust-request-chat-template \
226
+ --default-chat-template-kwargs '{"enable_thinking": false}' \
227
+ ...
228
+ ```
229
+
230
+ **Thinking ON for all requests (default if flag is omitted)**
231
+ ```bash
232
+ vllm serve MultiverseComputingCAI/Pulsar-16B-FP8 \
233
+ --served-model-name model \
234
+ --reasoning-parser-plugin nano_v3_reasoning_parser.py \
235
+ --reasoning-parser nano_v3 \
236
+ --trust-request-chat-template \
237
+ --default-chat-template-kwargs '{"enable_thinking": true}' \
238
+ ...
239
+ ```
240
+
241
+
242
+ ---
243
+
244
+ #### Per-request override
245
+
246
+ > **`--trust-request-chat-template` is required** to allow per-request overrides.
247
+
248
+ Individual requests can override the server default by passing `chat_template_kwargs` in the request body. This works regardless of the server-level default.
249
+
250
+ **Thinking ON/OFF for one request**
251
+ ```python
252
+ import requests
253
+
254
+ response = requests.post("http://localhost:8000/v1/chat/completions", json={
255
+ "model": "model",
256
+ "messages": [{"role": "user", "content": "Solve: x² - 5x + 6 = 0"}],
257
+ "max_tokens": 1024,
258
+ "temperature": 1.0,
259
+ "chat_template_kwargs": {"enable_thinking": True},
260
+ })
261
+ ```
262
+
263
+ ---
264
+
265
+ ## Tool Calling
266
+
267
+ Pulsar emits tool calls in the following format:
268
+
269
+ ```
270
+ <tool_call>
271
+ <function=get_weather>
272
+ <parameter=city>Paris</parameter>
273
+ <parameter=unit>celsius</parameter>
274
+ </function>
275
+ </tool_call>
276
+ ```
277
+
278
+ When serving (e.g with vLLM), you **must** use the `qwen3_coder` tool parser.
279
+
280
+ ```bash
281
+ vllm serve <model_path> \
282
+ --enable-auto-tool-choice \
283
+ --tool-call-parser qwen3_coder \
284
+ --trust-remote-code
285
+ ```
286
+
287
+ ## Training & Fine-Tuning
288
+
289
+ ### Base Model: NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
290
+
291
+ The base model [nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16) is a large language model (LLM) trained from scratch by NVIDIA, and designed as a unified model for both reasoning and non-reasoning tasks. The model's reasoning capabilities can be configured through a flag in the chat template. See the [original model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16) for details.
292
+
293
+
294
+ ### CompactifAI Compression
295
+
296
+ CompactifAI was applied to produce a smaller, efficient model (16B parameters) while aiming to preserve reasoning and tool-use capabilities. Supervised Fine Tuning was applied for improving cabapilities.
297
+
298
+ ### FP8 Post-Training Quantization
299
+
300
+ After compression and fine-tuning, the BF16 checkpoint was quantized to FP8 using [NVIDIA TensorRT Model Optimizer](https://github.com/NVIDIA/TensorRT-Model-Optimizer) (`modelopt` 0.42.0). A few selected modules (`lm_head`, all Mamba2 `conv1d` layers and the attention/Mamba projections of selected layers) were kept in higher precision.
301
+
302
+ ---
303
+
304
+ ## Evaluation & Benchmarks
305
+
306
+ ![Combined benchmark chart](assets/benchmarks.png)
307
+
308
+ | Benchmark | Pulsar-16B (BF16) | Pulsar-16B (FP8) | Pulsar-16B (NVFP4) |
309
+ |----------------|-------------------:|------------------:|--------------------:|
310
+ | SciCode | 25.84 | 23.08 | 23.28 |
311
+ | AA-LCR | 29.33 | 21.33 | 25.67 |
312
+ | IFBench | 70.79 | 69.60 | 69.90 |
313
+ | GPQA-D | 71.41 | 70.61 | 71.11 |
314
+ | AIME 2025 | 87.22 | 86.67 | 82.00 |
315
+ | MMLU-Pro | 74.78 | 74.76 | 74.19 |
316
+ | LiveCodeBench | 68.04 | 68.68 | 65.60 |
317
+ | Tau2-bench | 37.43 | 31.55 | 35.92 |
318
+
319
+ ### Performance
320
+ ![Performance results](assets/performance.png)
321
+ - **Framework:** [guidellm](https://github.com/vllm-project/guidellm)
322
+ - **Inference:** vLLM 0.18.0
323
+ - **Decode:** temperature=1.0, top_p=1.0
324
+ - **Phase duration:** Each phase lasts 3 minutes (excluding ramp-up and cool-down periods).
325
+ - **Workload shape:** Sweep across different levels of concurrency. 180s measure windows with ramp-up and cooldown excluded. 8k/16k workload as in the original model's card.
326
+ - **Estimated Peak Memory Usage:** KV cache utilization is monitored during the phase and we estimate memory usage as follows: `model_ weights_{gb} + kv_ cache_{usage_pct} × (nvml_used_{gb} − model_ weights_{gb})`
327
+
328
+ ### Long Context
329
+ ![Long-context benchmark results](assets/long_context_comparison.png)
330
+
331
+ Pulsar preserves strong long-context behavior after compression, tracking the Nemotron-3-Nano-30B-A3B baseline closely across retrieval-heavy and full-suite long-context evaluations. Results are reported for LongBench v1, AA-LCR, NIAH, and RULER groupings up to 256k context.
332
+
333
+ ### Evaluation Methodology
334
+
335
+ Benchmark scores were obtained with the following setups. Methodology varies by benchmark family.
336
+
337
+ #### MMLU-Pro, AIME25, GPQA:d, LiveCodeBench, IFBench, AA-LCR, SciCode
338
+
339
+ - **Framework**: [Nemo-Skills](https://github.com/NVIDIA-NeMo/Skills)
340
+ - **Inference**: vLLM 0.18.0
341
+ - **Decode**: temp 1.0, top_p 1.0
342
+
343
+ #### BFCL v4
344
+
345
+ - **Framework**: [EvalScope](https://github.com/EvalScope/EvalScope) 1.4.1
346
+ - **Inference**: vLLM 0.18.0
347
+ - **Decode**: temp 0.6, top_p 0.95, parallel_tool_calls true
348
+
349
+ #### Tau2-bench (Telecom)
350
+
351
+ - **Framework**: [EvalScope](https://github.com/EvalScope/EvalScope) 1.4.1
352
+ - **Inference**: vLLM 0.18.0
353
+ - **Decode (agent)**: temp 0.6, top_p 0.95
354
+ - **Decode (judge / user sim)**: temp 0.7, top_p 0.8, top_k 20, min_p 0
355
+ - **Repro**: telecom subset; max steps 100; 3 runs; judge parser hermes
356
+
357
+ #### LongBench v1
358
+
359
+ - **Framework**: CompactifAI evaluation interface with the LongBench harness, benchmark version v1, full suite.
360
+ - **Dataset**: `Xnhyacinth/LongBench`, all LongBench v1 tasks, no subset limit.
361
+ - **Inference**: vLLM 0.18.0
362
+ - **Decode**: max_tokens 1024, temperature 1.0, top_p 0.95, chat template enabled.
363
+
364
+ #### AA-LCR
365
+
366
+ - **Framework**: [EvalScope](https://github.com/EvalScope/EvalScope) 1.4.1
367
+ - **Dataset**: `evalscope/AA-LCR` from ModelScope, default test split, 300 examples, 3 repeats.
368
+ - **Inference**: vLLM 0.18.0
369
+ - **Decode**: temperature 1.0, top_p 1.0
370
+ - **Decode (judge)**: judge top_p 0.8, top_k 20, min_p 0.0; pattern scoring.
371
+
372
+ #### NIAH
373
+
374
+ - **Framework**: [Nemo-Skills](https://github.com/NVIDIA-NeMo/Skills)
375
+ - **Dataset / tasks**: RULER NIAH variants: single needle, multi-key, multi-query, and multi-value.
376
+ - **Context lengths**: 4096, 8192, 16384, 32768, 65536, 131072, and 262144 tokens.
377
+ - **Inference**: vLLM 0.18.0
378
+ - **Decode**: max_tokens 128, temperature 1.0, top_p 1.0, top_k 20, min_p 0.0, tokenizer `meta-llama/Llama-3.1-8B`.
379
+
380
+ #### RULER
381
+
382
+ - **Framework**: [Nemo-Skills](https://github.com/NVIDIA-NeMo/Skills)
383
+ - **Dataset / tasks**: RULER task families including VT, CWE, FWE, QA, and NIAH; CWE included.
384
+ - **Context lengths**: 4096, 8192, 16384, 32768, 65536, 131072, and 262144 tokens.
385
+ - **Inference**: vLLM 0.18.0
386
+ - **Decode**: temperature 1.0, top_p 1.0, top_k 20, min_p 0.0; task-specific output limits: NIAH 128, VT 30, CWE 120, FWE 50, QA 32.
387
+
388
+ ---
389
+
390
+ ## Languages
391
+
392
+ - **Primary language**: English
393
+ - **Other languages**: Spanish
394
+
395
+ Trained mainly on English with added Spanish. No systematic evaluation for languages outside English and Spanish.
396
+
397
+
398
+
399
+ ## Safety & Limitations
400
+
401
+ ### Known Limitations
402
+
403
+ - English-centric training data (inherited from base model).
404
+ - Tool calling depends on correct schema and tool design; exact parity with the original model is not guaranteed.
405
+ - Compression and quantization may affect some behaviors; evaluate for your use case.
406
+
407
+ ### Recommendations
408
+
409
+ - Validate tool outputs before running them
410
+ - Human oversight for critical use
411
+ - Task-specific eval before production
412
+
413
+ ---
414
+
415
+ ## Model Information
416
+
417
+ | Field | Value |
418
+ |--------------|--------------------- |
419
+ | Model name | Pulsar 16B FP8 |
420
+ | Based on | [MultiverseComputingCAI/Pulsar-16B](https://huggingface.co/MultiverseComputingCAI/Pulsar-16B) (originally [NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16)) |
421
+ | Precision | FP8 weights & activations, FP8 KV cache |
422
+ | Quantization | NVIDIA TensorRT Model Optimizer (`modelopt` 0.42.0) |
423
+ | Version | v1.5.0 |
424
+ | Release date | TBD |
425
+ | Developed by | Multiverse Computing |
426
+ | License | Apache 2.0 |
427
+ | Contact | business@multiversecomputing.com |
428
+
429
+ ---
430
+
431
+ ## Citation
432
+
433
+ If you use this model, please cite the base model and pulsar:
434
+
435
+ ```bibtex
436
+ @misc{nemotron3nanoTR,
437
+ title = {NVIDIA Nemotron 3 Nano Technical Report},
438
+ author = {{NVIDIA}},
439
+ year = {2025},
440
+ url = {https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Nano-Technical-Report.pdf}
441
+ }
442
+ @misc{pulsar16b,
443
+ title = {Pulsar 16B: Model developed from NVIDIA Nemotron-3-Nano-30B-A3B},
444
+ author = {Multiverse Computing},
445
+ year = {2026},
446
+ url = {https://huggingface.co/MultiverseComputingCAI/Pulsar-16B},
447
+ note = {Model developed based on nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 using CompactifAI technology}
448
+ }
449
+ @misc{pulsar16bfp8,
450
+ title = {Pulsar 16B FP8: FP8 post-training-quantized release of Pulsar-16B},
451
+ author = {Multiverse Computing},
452
+ year = {2026},
453
+ url = {https://huggingface.co/MultiverseComputingCAI/Pulsar-16B-FP8},
454
+ note = {FP8 quantization produced with NVIDIA TensorRT Model Optimizer}
455
+ }
456
+ ```
457
+
458
+ **Built by [Multiverse Computing](https://www.multiversecomputing.com)** · [Report an issue](TODO_PULSAR_HF_URL/discussions) · [Discord](https://discord.gg/cGas9uStqp)