feat(gemma4e-dflash-flash): nemotron3.5-bolt-dflash-flash:e4b (single-turn flash + DFlash speculation)

.github/script/ci/changes.py

@@ -82,10 +82,11 @@
     ("qwen3_8mtp", "qwen3_8_mtp"):           ("qwen3.8-mtp",),
     ("gemma", "gemma3"):                     ("gemma3",),
     ("gemma_text", "gemma3_text"):           ("gemma3-text",),
-    # Gemma4e_Flash and Gemma4e_DFlash derive from Gemma4e and their engines include gemma4e_npu.
-    ("gemma4e",):                            ("gemma4e", "gemma4e-flash", "gemma4e-dflash"),
+    # Gemma4e_Flash, Gemma4e_DFlash and Gemma4e_DFlash_Flash derive from Gemma4e and their engines include gemma4e_npu.
+    ("gemma4e",):                            ("gemma4e", "gemma4e-flash", "gemma4e-dflash", "gemma4e-dflash-flash"),
     ("gemma4e_flash",):                      ("gemma4e-flash",),
     ("gemma4e_dflash",):                     ("gemma4e-dflash",),
+    ("gemma4e_dflash_flash",):               ("gemma4e-dflash-flash",),
     ("gemma4_12b",):                         ("gemma4-12b",),
     ("hunyuan",):                            ("hunyuan",),
     ("gpt_oss",):                            ("gpt-oss",),

.github/script/npu_test/plan.py

@@ -33,6 +33,7 @@
     "gemma4e": "gemma4-it:e2b",
     "gemma4-12b": "gemma4-it:12b",
     "gemma4e-flash": "gemma4e-flash:e2b",
+    "gemma4e-dflash-flash": None,   # same 9 GB nemotron3.5-bolt:e4b package; runs when the change touches it
     "gemma4e-dflash": None,   # needs the 9 GB nemotron3.5-bolt:e4b package; runs when the change touches it
     "qwen3.5": "qwen3.5:0.8b",
     "minicpm-v-4.7": "minicpm-v-4.7:1b",

docs/docs/benchmarks/flash_results.md

@@ -31,6 +31,7 @@ AMD Ryzen™ AI 9 370 (Strix Point) with 32 GB DRAM; performance is comparable t
 |-------------------------|-----------|-------:|--------:|--------:|--------:|---------:|
 | **gemma4e-flash:e2b**   | XDNA2 | 0.20 | 0.22 | 0.36 | 0.73 | 1.53 |
 | **gemma4e-flash:e4b**   | XDNA2 | 0.33 | 0.34 | 0.62 | 1.20 | 2.57 |
+| **nemotron3.5-bolt-dflash-flash:e4b** | XDNA2 | 0.32 | 0.33 | 0.60 | 1.18 | 2.53 |
 | **qwen3vl-flash:4b**    | XDNA2 | 0.24 | 0.24 | 0.45 | 0.85 | 1.83 |
 | **hy-mt2-flash:1.8b**   | XDNA2 | 0.13 | 0.14 | 0.23 | 0.43 | 0.84 |
 
@@ -44,6 +45,7 @@ Prompt tokens divided by TTFT (TTFT includes generating the first token).
 |-------------------------|-----------|-------:|--------:|--------:|--------:|---------:|
 | **gemma4e-flash:e2b**   | XDNA2 | 322 | 593 | 721 | 705 | 668 |
 | **gemma4e-flash:e4b**   | XDNA2 | 193 | 373 | 412 | 427 | 398 |
+| **nemotron3.5-bolt-dflash-flash:e4b** | XDNA2 | 201 | 392 | 427 | 434 | 405 |
 | **qwen3vl-flash:4b**    | XDNA2 | 271 | 526 | 570 | 601 | 558 |
 | **hy-mt2-flash:1.8b**   | XDNA2 | 490 | 923 | 1,093 | 1,197 | 1,214 |
 
@@ -55,7 +57,10 @@ Prompt tokens divided by TTFT (TTFT includes generating the first token).
 |-------------------------|-----------|-------:|------------|
 | **gemma4e-flash:e2b**   | XDNA2 | 22.6 | [Gemma 4](../gemma4_results/) (`gemma4-it:e2b`) |
 | **gemma4e-flash:e4b**   | XDNA2 | 12.6 | [Gemma 4](../gemma4_results/) (`gemma4-it:e4b`) |
+| **nemotron3.5-bolt-dflash-flash:e4b** | XDNA2 | 18.2 | [Nemotron Bolt DFlash](../nemotron3.5-bolt-dflash_results/) (`nemotron3.5-bolt-dflash:e4b`) |
 | **qwen3vl-flash:4b**    | XDNA2 | 19.6 | [Qwen3](../qwen3_results/) (Qwen 3 4B) |
 | **hy-mt2-flash:1.8b**   | XDNA2 | 52.2 | [Hy-MT2](../hy-mt2_results/) |
 
 > Context is fixed at 1k tokens for every flash model, so 1k is the only applicable decode starting point.
+>
+> `nemotron3.5-bolt-dflash-flash:e4b` decodes with speculation, so its speed depends on the text: the value above is `flm bench` (about 30 decoded tokens, mean of 3 runs, 15 to 24 tokens/s between runs). Code and structured answers reach 33 to 40 tokens/s, open prose about 20 (see its [model card](../../models/nemotron3.5-bolt-dflash/#-model-card-nvidia-nemotron-35-bolt-e4b--dflash-flash)). The three rows of this model were measured on an AMD Ryzen™ AI MAX+ PRO 395 (Strix Halo) with the same method; on that machine `gemma4e-flash:e4b` gives 0.32 / 0.32 / 0.58 / 1.17 / 2.52 s.

docs/docs/instructions/server/openapi.md

@@ -507,7 +507,7 @@ gc.collect()
 
 ## ⚡ Example: Flash Models (Single-Turn, Pinned System Prompt)
 
-**Flash** models — `gemma4e-flash:e2b`, `gemma4e-flash:e4b`, `qwen3vl-flash:4b`, and `hy-mt2-flash:1.8b` — trade context length for a much faster time-to-first-token. They are built for high-throughput, short, independent requests: classification, captioning, extraction, routing, translation.
+**Flash** models — `gemma4e-flash:e2b`, `gemma4e-flash:e4b`, `nemotron3.5-bolt-dflash-flash:e4b`, `qwen3vl-flash:4b`, and `hy-mt2-flash:1.8b` — trade context length for a much faster time-to-first-token. They are built for high-throughput, short, independent requests: classification, captioning, extraction, routing, translation.
 
 They are **single-turn**: every request starts from a clean KV state, so earlier turns are not carried over. The system prompt is the one exception — it is prefilled once and cached, so keep it fixed and vary only the user message (and, for vision models, the image).
 
@@ -571,6 +571,6 @@ gc.collect()
 > 📏 Context is **fixed at 1k tokens** for every flash model; context-length overrides are ignored.  
 > 🔧 **Tool calling is not supported** — any `tools` you pass are dropped.  
 > 🌐 `hy-mt2-flash` is text-only and has no default system prompt — pin your own translation instruction as the system prompt, or put it in the user message.  
-> 🖼️ `gemma4e-flash` defaults to the smallest visual token budget (70) and keeps only the **first 30 seconds** of an audio clip. `qwen3vl-flash` always downscales images to a **256 px longer side**.  
+> 🖼️ `gemma4e-flash` and `nemotron3.5-bolt-dflash-flash` default to the smallest visual token budget (70) and keeps only the **first 30 seconds** of an audio clip. `qwen3vl-flash` always downscales images to a **256 px longer side**.  
 
-For the full per-model details, see the [`gemma4e-flash:e2b`](https://fastflowlm.com/docs/models/gemma/#-model-card-gemma-4-e2b-it--flash), [`gemma4e-flash:e4b`](https://fastflowlm.com/docs/models/gemma/#-model-card-gemma-4-e4b-it--flash), [`qwen3vl-flash:4b`](https://fastflowlm.com/docs/models/qwen/#-model-card-qwen3-vl-4b-instruct--flash), and [`hy-mt2-flash:1.8b`](https://fastflowlm.com/docs/models/hy-mt2/) model cards. Performance numbers are on the [Flash benchmarks](https://fastflowlm.com/docs/benchmarks/flash_results/) page.
+For the full per-model details, see the [`gemma4e-flash:e2b`](https://fastflowlm.com/docs/models/gemma/#-model-card-gemma-4-e2b-it--flash), [`gemma4e-flash:e4b`](https://fastflowlm.com/docs/models/gemma/#-model-card-gemma-4-e4b-it--flash), [`nemotron3.5-bolt-dflash-flash:e4b`](https://fastflowlm.com/docs/models/nemotron3.5-bolt-dflash/#-model-card-nvidia-nemotron-35-bolt-e4b--dflash-flash), [`qwen3vl-flash:4b`](https://fastflowlm.com/docs/models/qwen/#-model-card-qwen3-vl-4b-instruct--flash), and [`hy-mt2-flash:1.8b`](https://fastflowlm.com/docs/models/hy-mt2/) model cards. Performance numbers are on the [Flash benchmarks](https://fastflowlm.com/docs/benchmarks/flash_results/) page.

docs/docs/instructions/server/prefix_cache.md

@@ -41,7 +41,7 @@ The other supported models gain the same way. Measured on shorter prompts (about
 | Endpoint | `POST /v1/chat/completions` of `flm serve` (what Hermes, Open WebUI and most agents use), streaming or not |
 | Not covered | `flm run` (CLI mode) and the Ollama endpoints `/api/chat` and `/api/generate`, which keep the older single-checkpoint reuse. A request on an Ollama endpoint also empties the prefix cache, so a client that mixes both APIs loses its cached prompts. |
 
-Not covered: the single-turn models (`hy-mt2-flash`, `qwen3vl-flash`, `gemma4e-flash`: they take one request and pin their system prompt, so there is nothing to reuse), the developer-preview `*-core` backends, `lfm3`, and the speech, embedding and robotics models. They are unchanged.
+Not covered: the single-turn models (`hy-mt2-flash`, `qwen3vl-flash`, `gemma4e-flash`, `nemotron3.5-bolt-dflash-flash`: they take one request and pin their system prompt, so there is nothing to reuse), the developer-preview `*-core` backends, `lfm3`, and the speech, embedding and robotics models. They are unchanged.
 
 It is on by default. Nothing to configure for a single chat or agent.
 

docs/docs/instructions/server/request_parameters.md

@@ -70,7 +70,7 @@ A request that sends no sampling fields uses the model's own defaults, which dif
 | Models | `presence_penalty` at load |
 |---|---|
 | `qwen3.5:*`, `qwen3.5-core:*`, `qwen3.6-moe:35b-a3b`, `qwen3.6-moe-core:35b-a3b`, `qwen3.8-mtp:27b`, `minicpm-v-4.7:1b` | 1.5 |
-| `gemma4-it:*`, `gemma4-it-core:*`, `gemma4e-flash:*`, `nemotron3.5-bolt:e4b` | 1.0 |
+| `gemma4-it:*`, `gemma4-it-core:*`, `gemma4e-flash:*`, `nemotron3.5-bolt:e4b`, `nemotron3.5-bolt-dflash:e4b`, `nemotron3.5-bolt-dflash-flash:e4b` | 1.0 |
 | every other model | 0 |
 
 Sending `presence_penalty=0` turns the default off, and it stays off until the model is reloaded (see *Settings stick* above).
@@ -260,7 +260,7 @@ Before generating, FLM reads the prompt into the model's KV cache. This is the *
 
 | Models | `max_prefill_len` |
 |---|---|
-| Flash models (`gemma4e-flash:*`, `qwen3vl-flash:4b`, `hy-mt2-flash:1.8b`) | 1024 |
+| Flash models (`gemma4e-flash:*`, `nemotron3.5-bolt-dflash-flash:e4b`, `qwen3vl-flash:4b`, `hy-mt2-flash:1.8b`) | 1024 |
 | Every other chat model | 4096 |
 
 **`max_prefill_len` is not a limit on the prompt.** A longer prompt just takes more chunks. A 32,000-token prompt on a model with 4096-token chunks is read in eight passes, seven of 4096 tokens and one of 3,328, and the server log shows each one:

docs/docs/models/flash.md

@@ -13,6 +13,7 @@ parent: Models
 |---------------|----------|----------------|
 | `gemma4e-flash:e2b` | Any-to-Text (image, audio) | [Gemma 4 E2B — Flash](https://fastflowlm.com/docs/models/gemma/#-model-card-gemma-4-e2b-it--flash) |
 | `gemma4e-flash:e4b` | Any-to-Text (image, audio) | [Gemma 4 E4B — Flash](https://fastflowlm.com/docs/models/gemma/#-model-card-gemma-4-e4b-it--flash) |
+| `nemotron3.5-bolt-dflash-flash:e4b` | Any-to-Text (image, audio), speculative decoding | [Nemotron Bolt E4B — DFlash Flash](https://fastflowlm.com/docs/models/nemotron3.5-bolt-dflash/#-model-card-nvidia-nemotron-35-bolt-e4b--dflash-flash) |
 | `qwen3vl-flash:4b` | Image-Text-to-Text | [Qwen3-VL-4B — Flash](https://fastflowlm.com/docs/models/qwen/#-model-card-qwen3-vl-4b-instruct--flash) |
 | `hy-mt2-flash:1.8b` | Text-to-Text (Translation) | [Hy-MT2](https://fastflowlm.com/docs/models/hy-mt2/) |
 

docs/docs/models/nemotron3.5-bolt-dflash.md

@@ -58,3 +58,61 @@ Expected: `1, 2, 3, ..., 40` followed by `done`, at a visibly higher speed than
 - Speculation pauses while a tool-call or JSON grammar constrains the output, and whenever it would not be faster (for example at very
   long contexts); it resumes by itself.
 - `flm remove nemotron3.5-bolt-dflash:e4b` removes only the draft package; `nemotron3.5-bolt:e4b` stays installed.
+
+---
+
+## 🧩 Model Card: [NVIDIA-Nemotron-3.5-Bolt-E4B — DFlash Flash](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB)
+
+- **Type:** Any-to-Text (image, audio), single-turn, with speculative decoding
+- **What it is:** [`nemotron3.5-bolt-dflash:e4b`](#-model-card-nvidia-nemotron-35-bolt-e4b-with-the-dflash-draft) built for the
+  [Flash](https://fastflowlm.com/docs/models/flash/) serving style: the fast short-prompt prefill of `gemma4e-flash:e4b` together with
+  DFlash speculative decoding. Every request starts from a clean state; the system prompt is pinned and reused
+  ([how flash models work in server mode](https://fastflowlm.com/docs/instructions/server/openapi/#-example-flash-models-single-turn-pinned-system-prompt))
+- **Time to first token:** the same as `gemma4e-flash:e4b`, about 0.3 s for a short prompt, against 1.2 s for `nemotron3.5-bolt-dflash:e4b`
+- **Decode speed:** 33 to 40 tokens/s on code and structured answers, 19 to 21 tokens/s on open prose, against 13 for `gemma4e-flash:e4b`
+- **Think / Conversational mode / Vision / Audio:** as `nemotron3.5-bolt:e4b`. Tool calling is not supported (the `tools` field is dropped).
+  Images default to the smallest visual token budget (70) and only the first 30 seconds of an audio clip are used
+- **Packages:** the same two downloads as `nemotron3.5-bolt-dflash:e4b` (`Nemotron-3.5-Bolt-E4B-NPU2`, about 9 GB, and the 0.27 GB draft
+  `Nemotron-3.5-Bolt-E4B-DFlash-NPU2`). Nothing new to download if you already have that tag
+- **Max / Default Context Length:** 1k tokens (fixed, like every flash model)
+- **License:** [OpenMDW-1.1](https://openmdw.ai/license/1-1/)
+- **FastFlowLM:** 1.0.13 or later
+
+▶️ Run with FastFlowLM in PowerShell:
+
+```shell
+flm run nemotron3.5-bolt-dflash-flash:e4b
+```
+
+Server mode:
+
+```shell
+flm serve nemotron3.5-bolt-dflash-flash:e4b
+```
+
+📊 **Measured** (greedy, 18-token pinned system prompt, NPU, quiet machine):
+
+| Prompt length | `gemma4e-flash:e4b` | `nemotron3.5-bolt-dflash:e4b` | `nemotron3.5-bolt-dflash-flash:e4b` |
+|---|---|---|---|
+| ~50 tokens, time to first token | 0.31 s | 1.24 s | 0.32 s |
+| ~140 tokens | 0.57 s | 1.24 s | 0.60 s |
+| ~540 tokens | 1.48 s | 1.84 s | 1.49 s |
+| ~960 tokens | 2.51 s | 2.15 s | 2.52 s |
+| decode, code answer | 13 tokens/s | 38 to 40 tokens/s | 33 to 40 tokens/s |
+| decode, prose answer | 13 tokens/s | 17 to 22 tokens/s | 19 to 21 tokens/s |
+
+📝 **Note:**
+
+- Flash prefill is tuned for short prompts: it is the faster choice up to a prompt of roughly 700 tokens. For longer prompts use
+  `nemotron3.5-bolt-dflash:e4b`, which also has a longer context.
+- A long, fixed system prompt (instructions, examples) is where this tag gains most: it is prefilled once and the draft model sees it on
+  every request, which keeps the speculative steps long. Changing the system prompt between requests discards the pin.
+- Sampling, `FLM_DFLASH=0` and the other behaviours of speculation are as on the [`nemotron3.5-bolt-dflash:e4b`](#-model-card-nvidia-nemotron-35-bolt-e4b-with-the-dflash-draft) card.
+  Greedy output can differ from `nemotron3.5-bolt:e4b` at near-ties.
+- Accuracy (50-question samples, XRT and HRX give the same results): GSM8K 92%, the same as `nemotron3.5-bolt-dflash:e4b` and
+  `gemma4e-flash:e4b`. Document reading (DocVQA, ANLS) is 45, against 76 for `nemotron3.5-bolt-dflash:e4b`: flash models look at an image
+  with the smallest visual token budget (70 instead of 280), which is enough for captions and simple scenes but not for small text. For
+  documents and screenshots use `nemotron3.5-bolt-dflash:e4b`.
+- Needs 9 of the NPU's 16 hardware contexts (vision and audio included), so other models can still be loaded next to it.
+- `flm remove nemotron3.5-bolt-dflash-flash:e4b` removes only the draft package, shared with `nemotron3.5-bolt-dflash:e4b`: remove
+  both tags' use of it first if you want the space back.

src/CMakeLists.txt

@@ -514,6 +514,7 @@ list(APPEND FLM_ENGINE_LINK_LIBS
 
 list(APPEND FLM_ENGINE_LINK_LIBS gemma4e_flash)
 list(APPEND FLM_ENGINE_LINK_LIBS gemma4e_dflash)
+list(APPEND FLM_ENGINE_LINK_LIBS gemma4e_dflash_flash)
 
 list(APPEND FLM_ENGINE_LINK_LIBS
     gemma4_12b_npu

src/common/AutoModel/builtin_backends.cpp

@@ -54,6 +54,7 @@ void register_builtin_backends(BackendRegistry& registry) {
     RegisterFlm<gemma4e_npu>(registry, "gemma4e");
     RegisterFlm<gemma4e_flash>(registry, "gemma4e-flash");
     RegisterFlm<gemma4e_dflash>(registry, "gemma4e-dflash");
+    RegisterFlm<gemma4e_dflash_flash>(registry, "gemma4e-dflash-flash");
     RegisterFlm<gemma4_12b_npu>(registry, "gemma4-12b");
     RegisterFlm<hunyuan_npu>(registry, "hunyuan");
     RegisterFlm<gpt_oss_npu>(registry, "gpt-oss");

src/common/AutoModel/modeling_gemma4e.cpp

@@ -704,7 +704,9 @@ bool Gemma4e::insert(chat_meta_info_t& meta_info, lm_uniform_input_t& input, std
         ? this->_shared_insert(meta_info, tokens, is_cancelled, &multi_modal_payload, last_image_token_index)
         : this->_shared_insert(meta_info, tokens, is_cancelled, nullptr);
 
-    if (this->enable_think) {
+    // Single-turn (flash): the only checkpoint is the pinned system prefix, which Gemma4e_Flash owns. Taking this one would
+    // overwrite it with the whole first prompt, and every later request would restore into that stale context.
+    if (this->enable_think && !this->single_turn) {
         checkpoint_his = token_history;
         int checkpoint_idx = this->lm_engine->checkpoint();
     }