SparkLLM 社区反响

核验提交：ba7074235b88 · 仓库更新时间：2026-11-01T04:28:09Z

• 源码证据：inputs/research_task.json

  "model_id": "SparkLLM/Spark-X2.6-4B",

• 源码证据：main.py

  client = ollama.Client(host=host)

核验提交：56cbbfeb7c5c · 仓库更新时间：2026-10-22T05:36:39Z

• 源码证据：utils/spark_int4.py

  """Load Spark-X2.6's official mixed INT4/INT8 checkpoint on Ascend."""

• 源码证据：pyproject.toml

  spark_int4 = "utils.register_spark:register_int4"

• 源码证据：utils/spark_int4.py

  y = torch_npu.npu_weight_quant_batchmatmul(

核验提交：d9510cac2945 · 仓库更新时间：2026-10-23T07:27:16Z

• 源码证据：index.mjs

  id: "spark-x2.6-4b",

• 源码证据：index.mjs

  id: "spark-x2.5",

• 源码证据：index.mjs

  export const XfyunAuthPlugin = async (input, options) => {

核验提交：cbd92e877b6d · 仓库更新时间：2026-10-21T18:28:17Z

• 源码证据：Fatima_fellowship.ipynb

  model_id = "XHToken/Spark-X2.6-4B"

• 源码证据：README.md

  - 5 samples per condition, one task, one model.

核验提交：b9ba007ee4d2 · 仓库更新时间：2026-10-31T17:32:40Z

• 源码证据：src/rizzo_flow/config.py

  ModelSpec("4b", "XHToken/Spark-X2.5-4B", "0bcb35678590218655dff3765b9e61c83b35e9c4", 2560),

• 源码证据：src/rizzo_flow/backend_llama.py

  ARCHITECTURE = "spark2_5"

• 源码证据：training/train_lora.py

  config = LoraConfig(

核验提交：89284376c98a · 仓库更新时间：2026-11-02T01:29:26Z

• 源码证据：crates/runtime/src/families/spark/mod.rs

  //! The `spark2_5` decode flow (`XHToken/Spark-X2.5-4B`), the NINTH flow.

• 源码证据：crates/model-io/src/arch_baselines/spark.rs

  /// Canonical `Spark-X2.5-4B` baseline: 36 DENSE layers, GQA at 16 query

• 源码证据：crates/runtime/tests/real_forward_spark.rs

  let arch = build_synthetic_spark_install(&dir, VOCAB, LAYERS, "spark-toy")

• 源码证据：crates/lab/families/spark2_5.json

  "family": "spark2_5",

• 源码证据：crates/runtime/src/real_forward_types.rs

  | model_io::ModelFamily::Spark25

• 源码证据：crates/runtime/src/real_forward_types.rs

  | model_io::ModelFamily::Flint => gpu::DecodeAttentionKernel::GqaShared,

核验提交：168745734f22 · 仓库更新时间：2026-10-30T08:52:54Z

• 源码证据：app/src/main/java/com/druk/lmplayground/models/ModelInfoProvider.kt

  remoteUri = Uri.parse("https://huggingface.co/XHToken/Spark-X2.5-4B-GGUF/resolve/main/Spark-X2.5-4B-Q4_K_M.gguf"),

• 源码证据：app/src/main/cpp/LlamaModel.cpp

  model = llama_model_load_from_file(modelPath.c_str(), model_params);

• 源码证据：app/src/main/java/com/druk/lmplayground/conversation/ChatRewind.kt

  fun regenerate(messages: List<Message>, assistantIndex: Int): Plan? {

• 源码证据：app/src/main/java/com/druk/lmplayground/conversation/ConversationViewModel.kt

  fun branchFrom(assistantMessageId: Long) {

• 源码证据：app/src/main/java/com/druk/lmplayground/inference/ModelRuntime.kt

  suspend fun rewindSession(

• 源码证据：app/src/main/cpp/LlamaGenerationSession.cpp

  int LlamaGenerationSession::rewindToTurn(int turnIndex, const char *expectedText) {

核验提交：138a3138b0e9 · 仓库更新时间：2026-09-18T06:22:40Z

• 源码证据：Makefile

  MODEL ?= $(HOME)/models/Spark-X2.5-4B-Q4_K_M.gguf

• 源码证据：mingo.go

  func (tb *Toolbox) Call(ctx context.Context, call ToolCall, allowed []ToolSpec) string {

核验提交：8ed7d85dea3f · 仓库更新时间：2026-10-15T17:03:40Z

• 源码证据：configs/models.json

  "display_name": "Spark-X2.5-4B",

• 源码证据：results/current/spark-reference/capacity/manifest.json

  "engine_source": "XHToken/llama.cpp",

• 源码证据：bench/models.py

  def primary_arms():

核验提交：c983fa887c5a · 仓库更新时间：2026-10-25T23:00:17Z

• 源码证据：benchmark.json

  "hf.co/peculiar-ragdoll/Sharp-Spark-X2.5-4B-GGUF:Q4_K_XL-no_think": {

• 源码证据：inference.py

  def process_problem_files(problems_dir, template_content, endpoints: List[Endpoint], language, max_problem_number=9999,

• 源码证据：benchmark.json

  "hf.co/XHToken/Spark-X2.6-4B-GGUF:Q4_K_M-no_think": {

核验提交：48e6fe7c1877 · 仓库更新时间：2026-09-21T02:01:06Z

• 源码证据：src/cli.rs

  #[arg(long, default_value = "XHToken/Spark-X2.5-1.7B")]

• 源码证据：src/model.rs

  pub struct SparkConfig {

• 源码证据：src/model.rs

  pub fn forward(&self, xs: &Tensor) -> CandleResult<Tensor> {

核验提交：39d8082a9550 · 仓库更新时间：2026-09-02T18:29:36Z

• 源码证据：README.md

  hf download XHToken/Spark-X2.5-1.7B --revision bbd9e9fdd10746eb2b612737b2811a288563b9f0 --local-dir models/Spark-X2.5-1.7B

• 源码证据：run_benchmark.py

  model = AutoModelForCausalLM.from_pretrained(

核验提交：ad82cf42c7f3 · 仓库更新时间：2026-09-10T14:41:16Z

• 源码证据：deploy.py

  REPO_ID = str(_hf.get("repoId") or "XHToken/Spark-X2.5-4B")

• 源码证据：deploy.py

  self.model = AutoModelForCausalLM.from_pretrained(

• 源码证据：entry.py

  def main() -> int:

核验提交：90c9c4cfbaad · 仓库更新时间：2026-09-11T06:09:09Z

• 源码证据：run_eval.py

  from spark_mlx_llm import load

• 源码证据：README.md

  | 模型 | XHToken/Spark-X2.5-1.7B |

核验提交：10257659a1ec · 仓库更新时间：2026-09-04T06:14:19Z

• 源码证据：README.md

  # Spark-X2.5-4B 真实仓库审查：定位第二断点并闭环 vLLM 0.23 工具调用

• 源码证据：evaluation/acceptance_vllm023.py

  from vllm_spark2_5_plugin.spark2_5 import Spark2_5ForCausalLM

核验提交：cd4496b46d40 · 仓库更新时间：2026-10-14T10:00:07Z

• 源码证据：Spark-X2.5-Catgirl-Reasoning/training/configs/config.yaml

  path: <DataRoot>/msp/models/Spark-X2.5-4B

• 源码证据：Spark-X2.5-Catgirl-Reasoning/protocol/evaluate_coldstart.py

  def generate(model,tok,question,seed,args):

核验提交：134872ddf411 · 仓库更新时间：2026-09-12T13:00:26Z

• 源码证据：run_eval.py

  MODEL_ID = 'XHToken/Spark-X2.5-1.7B'

• 源码证据：run_eval.py

  model = AutoModelForCausalLM.from_pretrained(model_dir, local_files_only=True,

核验提交：785322d17f2d · 仓库更新时间：2026-11-02T01:31:17Z

• 源码证据：conversion/spark2_5.py

  @ModelBase.example("XHToken/Spark-X2.5-1.7B")

• 源码证据：src/models/spark2-5.cpp

  void llama_model_spark2_5::load_arch_tensors(llama_model_loader &) {

核验提交：680df5c670d7 · 仓库更新时间：2026-11-02T00:51:15Z

• 源码证据：swift/model/models/spark.py

  Model('XHToken/Spark-X2.5-4B'),

• 源码证据：swift/model/models/spark.py

  class Spark2_5Loader(ModelLoader):

• 源码证据：swift/template/templates/spark.py

  register_template(Spark2_5TemplateMeta(LLMTemplateType.spark2_5, template_cls=Spark2_5Template))

核验提交：5cec55e53bb1 · 仓库更新时间：2026-10-23T07:02:45Z

• 源码证据：spec_smoke/run_spec_smoke.py

  TARGET = os.path.join(os.path.dirname(ROOT), "Spark-X2.6-4B.Q8_0.gguf")

• 源码证据：spec_smoke/run_spec_smoke.py

  DRAFT = os.path.join(os.path.dirname(ROOT), "Spark-X2.6-4B.Q2_K.gguf")

• 源码证据：spec_smoke/analyze.py

  """Builds REPORT.md from the measured results.

核验提交：1c2862ce0064 · 仓库更新时间：2026-11-01T09:22:52Z

• 源码证据：harness/runner.py

  self.model.get("engine", self.cfg["engine"]),

• 源码证据：results/spark-x2.6-4b.json

  "score": 0.6057619408642911,

• 源码证据：results/spark-x2.5-4b.json

  "tok_per_s": 55.6,

核验提交：c239a048a8a0 · 仓库更新时间：2026-11-02T01:45:11Z

• 源码证据：src/shared/openai-compat-presets.ts

  defaultModel: 'spark-x2.5',

• 源码证据：src/services/worker/OpenAICompatProvider.ts

  const model = configuredModel || preset.defaultModel;

• 源码证据：docs/public/usage/openai-compatible-provider.mdx

  `iflytek` points at iFlytek's Astron MaaS platform and defaults to `spark-x2.5`.

• 源码证据：src/services/worker/OpenAICompatProvider.ts

  ...(withReasoning ? { reasoning: REASONING_DISABLED } : {}),

• 源码证据：tests/worker/reasoning-budget-exhaustion.test.ts

  it('openai-compatible: a strict server that rejects the reasoning field is retried without it', async () => {

• 源码证据：src/services/worker/streamed-chat-completion.ts

  failedBeforeOutput: true,

• 源码证据：src/services/worker/OpenAICompatProvider.ts

  const send = (withReasoning: boolean) => sendChatCompletion({

• 源码证据：src/services/worker/OpenAICompatProvider.ts

  protected override observationTurnPrompt(session: ActiveSession, first: PendingMessageWithId, prompt: ObservationPromptParts): string {

• 源码证据：src/services/worker/OpenAICompatProvider.ts

  ['bad_gateway', 502],

• 源码证据：src/services/worker/streamed-chat-completion.ts

  allowMaxCompletionTokensRetry?: boolean;

核验提交：f64eaf9f71c3 · 仓库更新时间：2026-10-27T18:41:28Z

• 源码证据：src/config.py

  model: str = "spark2.6-4B"

• 源码证据：src/llm_client.py

  "chat_template_kwargs": {"enable_thinking": False},

• 源码证据：src/usecase.py

  output = self._llm.complete(

• 源码证据：outputs/ej1-turno.md

  Si estás haciendo estudios con contraste, debes ir 8 horas antes de comer.

• 源码证据：src/validator.py

  def validate_output(output: str, profile: LanguageProfile) -> None:

• 源码证据：src/usecase.py

  validate_output(output, profile)

• 源码证据：src/ocr.py

  class TesseractOCR:

• 源码证据：src/tts.py

  class MacSayTTS:

• 源码证据：src/languages.py

  PROFILES: dict[str, LanguageProfile]

• 源码证据：src/validator.py

  def validate_numbers_preserved

核验提交：08d28f58a10d · 仓库更新时间：2026-10-23T16:37:45Z

• 源码证据：qualify_runtime.py

  MODEL = "XHToken/Spark-X2.6-4B"

• 源码证据：notice_reader.py

  self.model, self.tokenizer = load(model_path, dtype="bfloat16", tokenizer_config={"trust_remote_code": False})

• 源码证据：notice_reader.py

  raise IncompleteNoticeError("This notice is too long to finish. Try a shorter section.")

• 源码证据：evaluate_model.py

  "facts_to_check": ["Riverside Library", "Saturday 31 October"

核验提交：d69e3be75f05 · 仓库更新时间：2026-10-24T12:34:42Z

• 源码证据：engine.py

  DEFAULT_MODEL = Path(os.environ.get("RURAL_SIGNAL_MODEL", ROOT / "models" / "Spark-X2.6-4B-Q4_K_M.gguf"))

• 源码证据：engine.py

  completed = subprocess.run(command, text=True, capture_output=True, timeout=300, check=False)

• 源码证据：engine.py

  output = dict(case["locked"])

• 源码证据：app.py

  result = engine.run_case(CASES[request["case_id"]])

• 源码证据：README.md

  - The official `XHToken/Spark-X2.6-4B-GGUF`

核验提交：f4d4454b3b63 · 仓库更新时间：2026-10-24T18:33:10Z

• 源码证据：model_lock.json

  "repo": "XHToken/Spark-X2.6-4B-GGUF",

• 源码证据：model_lock.json

  "revision": "fe31f8d64890302cc3a1d365d700357b4b75ba77",

• 源码证据：server.py

  result = model.create_chat_completion(messages=messages,temperature=0,max_tokens=180,response_format={'type':'json_object'})

• 源码证据：planner.py

  def validate(raw):

• 源码证据：planner.py

  {'role':'system','content':'You organize a household preparedness checklist. Select the next up to three distinct task IDs from the supplied candidates. Prioritize unmet basic supplies and access or medication planning, considering the household facts and a use-what-I-have spending preference. Do not invent tasks, advice, quantities, contacts or facts. This is preparation before an emergency. Return only JSON with one key priorities containing an array of task IDs. Do not explain or think aloud.'},

核验提交：d07eb52a05c9 · 仓库更新时间：2026-10-25T11:54:33Z

• 源码证据：civicclear_colab.py

  MODEL_ID = "XHToken/Spark-X2.6-4B"

• 源码证据：civicclear_colab.py

  model = AutoModelForCausalLM.from_pretrained(MODEL_ID, device_map="auto", torch_dtype="auto", trust_remote_code=True)

• 源码证据：civicclear_colab.py

  output = model.generate(**inputs, max_new_tokens=700, do_sample=False)

• 源码证据：index.html

  tice.value.trim())return notice.focus();let p=plans[key];meaning.textContent=p.m;steps.innerHTML=list(p.s);docs.innerHTML=list(p.d);dates.innerHTML=list(p.t);risks.innerHTML=list(p.r);verify.textContent=p.v;results.classList.add('show');results.scrollIntoView({behavior:'smooth'})};</script></body></html>

• 源码证据：aegis_her_app.py

  MODEL=os.getenv("SPARK_SERVED_MODEL","SparkLLM/Spark-X2.6-4B")

• 源码证据：aegis_her_app.py

  def infer(clean_text,language,base=API_BASE,model=MODEL):

• 源码证据：aegis_her_app.py

  def validate(data):

• 源码证据：aegis_her_app.py

  REPLAY—NOT MODEL OUTPUT

核验提交：6088a60e71ae · 仓库更新时间：2026-11-01T03:33:20Z

• 源码证据：configs/paths.yaml

  base_model: /mnt/f/LOCAL-Models/XHToken/Spark-X2.6-4B

• 源码证据：configs/general_pilot.yaml

  # DRAFT (not run yet: Rishi reviews the architecture first).

• 源码证据：configs/general_pilot.yaml

  nfl_route: 0 # the NFL track is parked

• 源码证据：src/asktrainer/modelio.py

  pm = get_peft_model(model, cfg, autocast_adapter_dtype=True)

• 源码证据：src/asktrainer/train.py

  model = apply_lora(model, cfg.lora, cfg.optim.gradient_checkpointing)

• 源码证据：configs/general_full.yaml

  name: sft-main

• 源码证据：scripts/real_run.py

  def stalled(

• 源码证据：src/asktrainer/train.py

  if (run_dir.parent / PAUSE_FILE).exists():

• 源码证据：src/asktrainer/traceexport.py

  def gold(

• 源码证据：src/asktrainer/train.py

  def limit_cuda

• 源码证据：scripts/real_run.py

  def score_pending

• 源码证据：src/asktrainer/modelio.py

  thinking: bool = False,

• 源码证据：scripts/llamacpp/release_quants.ps1

  [string]$Name = 'Spark-X2.6-4B-SQL-Cypher-Agent',

核验提交：6a23516a614b · 仓库更新时间：2026-10-25T22:49:30Z

• 源码证据：inference/src/main/cpp/ai_chat.cpp

  // Direct-answer mode by default: GGUF templates such as Spark-X2.6 otherwise

• 源码证据：inference/src/main/cpp/ai_chat.cpp

  inputs.enable_thinking = g_enable_thinking;

• 源码证据：inference/src/main/cpp/ai_chat.cpp

  return common_chat_templates_apply(g_chat_templates.get(), inputs).prompt.str();

• 源码证据：tests/native/chat_template_thinking_test.cpp

  inputs.enable_thinking = thinking;

• 源码证据：app/src/main/java/me/ailinux/pocket/GgufLoadFailure.kt

  internal class GgufLoadFailure(

• 源码证据：app/src/main/java/me/ailinux/pocket/ModelStore.kt

  internal fun ensureNoSilentRangeRestart(offset: Long, code: Int)

• 源码证据：app/src/main/java/me/ailinux/pocket/MainActivity.kt

  GgufLoadFailure.Phase.NATIVE_MODEL_LOAD

核验提交：ccef76f5a28d · 仓库更新时间：2026-10-16T02:16:46Z

• 源码证据：apply.sh

  if [ -f "$DEST" ] && cmp -s "$SRC" "$DEST"; then

• 源码证据：apply.sh

  cp -p "$DEST" "$BAK"

• 源码证据：apply.sh

  cp "$SRC" "$DEST"

• 源码证据：patches/spark2_5.py

  class Spark2_5Model(nn.Module):

• 源码证据：verify.sh

  import mlx_lm.models.spark2_5 as sp

核验提交：b7b5bd81c256 · 仓库更新时间：2026-10-30T15:28:15Z

• 源码证据：recipe.toml

  model_repo = "XHToken/Spark-X2.6-4B"

• 源码证据：mac/engine.py

  def build_engine(

• 源码证据：mac/serve.py

  p = subprocess.Popen(

• 源码证据：mac/engine.py

  TENSORFOLD_REPO = "https://github.com/ashhart/TensorFold"

核验提交：9cc8db8c8f7e · 仓库更新时间：2026-10-25T17:30:11Z

• 源码证据：examples/spark-x2.6-4b/model.yaml

  repo: Spark-X2.6-4B-GGUF

• 源码证据：examples/spark-x2.6-4b/model.yaml

  variable: enable_thinking

• 源码证据：examples/spark-x2.6-4b/model.yaml

  variable: reasoning_effort

• 源码证据：lmstudio_reasoning.py

  def build_yaml(

核验提交：f48947d48e15 · 仓库更新时间：2026-10-27T02:31:46Z

• 源码证据：.env.example

  XFYUN_MODEL=spark-x2.6

• 源码证据：src/configService.js

  model: env.getEnv('XFYUN_MODEL') || '',

• 源码证据：src/xfyun.js

  model: cfg.model,

• 源码证据：src/xfyun.js

  const url = `${base}/v2/chat/completions`;

• 源码证据：src/configService.js

  showModelName: s.showModelName === undefined ? true : !!s.showModelName

• 源码证据：public/assets/app.js

  return uiSettings.showModelName === false ? c.name : `${c.name} · ${c.model}`;

• 源码证据：src/adminAuth.js

  function appendCookie(res, cookie) {

• 源码证据：src/adminAuth.js

  function clearAllAuthCookies(res) {

• 源码证据：src/routes/admin.js

  adminAuth.clearAllAuthCookies(res);

• 源码证据：src/shareService.js

  function resolveUploadPath(url) {

• 源码证据：public/assets/topbar-more.js

  function bind()

核验提交：1716901abe71 · 仓库更新时间：2026-10-25T15:30:01Z

• 源码证据：backend/pianke/config.py

  Spark-X2.5-4B

• 源码证据：backend/pianke/config.py

  Spark-X2.5-1.7B

• 源码证据：backend/pianke/llm/client.py

  r = httpx.post(f"{runtime.url(model)}/v1/chat/completions"

• 源码证据：backend/pianke/flows/recommend.py

  client.chat(

核验提交：33ac96a5ff61 · 仓库更新时间：2026-10-28T03:59:23Z

• 源码证据：moss_tts.c

  "spark-x2.6"

• 源码证据：moss_tts.c

  response.output_text.delta

• 源码证据：moss_tts.c

  /v1/responses

• 源码证据：moss_tts.c

  speak_text2(

核验提交：93349c886e12 · 仓库更新时间：2026-10-26T17:19:03Z

• 源码证据：agent/common.py

  spark-x2.6-4b-lossless

• 源码证据：agent/base_agent.py

  def ask(self, user_text):

• 源码证据：agent/base_agent.py

  if tool == "query_metric":

• 源码证据：agent/query_metric.py

  def tool_query_metric(

• 源码证据：data/add_views.py

  CREATE VIEW v_orders AS

• 源码证据：data/add_views.py

  CREATE VIEW v_customers AS

• 源码证据：data/add_views.py

  CREATE VIEW v_product_sales AS

• 源码证据：agent/common.py

  def load(self, sql, state):

• 源码证据：agent/common.py

  task["_ns_file"] = args["_ns_file"]

• 源码证据：agent/base_agent.py

  if tool == "load_df" and self.sandbox:

• 源码证据：agent/v13b_hybrid2.py

  self._path = self._choose_path(user_text)

• 源码证据：agent/v13c_gated.py

  if t["function"]["name"] in ("load_df", "run_python", "calculator", "query_metric")

• 源码证据：agent/v13d_forced.py

  self._path = "PANDAS"

• 源码证据：probe/probe.py

  def step_models(self):

• 源码证据：agent/common.py

  def register_tool(name, fn):

• 源码证据：agent/base_agent.py

  return EXTRA_TOOLS[tool](args, {"session_id": self.session_id, "trace": self.trace})

• 源码证据：agent/domain_hospital.py

  def hosp_metric(args, state=None):

• 源码证据：agent/domain_hospital.py

  'data', 'hospital.db')).execute(sql, params).fetchall()]

• 源码证据：agent/domain_hospital.py

  _register("hosp_metric", hosp_metric)

• 源码证据：agent/v12h_hospital.py

  class AgentV12H(AgentV9):

• 源码证据：agent/v9h_baseline.py

  class AgentV9H(AgentV9):

• 源码证据：data/add_views_hospital.py

  CREATE VIEW v_rx AS

• 源码证据：data/add_views_hospital.py

  CREATE VIEW v_visits AS

• 源码证据：data/add_views_hospital.py

  CREATE VIEW v_patients AS

• 源码证据：data/add_views_hospital.py

  CREATE VIEW v_drugs AS

• 源码证据：data/make_hospital.py

  np.random.seed(42)

• 源码证据：eval/build_cases_hospital.py

  TOOL_CASES = [

• 源码证据：eval/runner.py

  ap.add_argument("--cases", default="",

• 源码证据：agent/v12c_calcpy.py

  class AgentV12C(AgentV9):

核验提交：3e5a3ede2dfa · 仓库更新时间：2026-11-01T22:59:16Z

• 源码证据：README.md

  **Spark-X2.6-4B-SLQ-4.50bpw**

• 源码证据：README.md

  mlx_lm.load("hermitdave/Spark-X2.6-4B-SLQ-4.50bpw"

• 源码证据：src/slq_mlx/mlx_model_utils.py

  excluded = {"g_proj"}

• 源码证据：src/slq_mlx/mlx_model_utils.py

  projs = "|".join(sorted(p for p in proj_attn if p not in excluded))

• 源码证据：src/slq_mlx/scripts/run_gptq.py

  model, tokenizer = load_model(args.model, trust_remote_code=args.trust_remote_code, lazy=args.lazy)

• 源码证据：src/slq_mlx/mlx_quantizer.py

  def quantize(self, bitwidth_options: List[int], calibration_bitwidth: int):

• 源码证据：src/slq_mlx/mlx_pack.py

  def pack_checkpoint(

• 源码证据：src/slq_mlx/scripts/run_pack_mlx.py

  description="Export an SLQ store as an MLX-native packed checkpoint"

• 源码证据：pyproject.toml

  license = {text = "Apache-2.0"}

• 源码证据：src/slq_mlx/mlx_pack.py

  if scale_dtype == "float16":

• 源码证据：src/slq_mlx/mlx_metrics.py

  mx.clear_cache()

• 源码证据：src/slq_mlx/mlx_pack.py

  def pack_checkpoint_streaming(

• 源码证据：src/slq_mlx/scripts/run_pack_mlx.py

  if args.streaming:

核验提交：f92a2a83d454 · 仓库更新时间：2026-11-01T18:57:21Z

• 源码证据：xinference/model/llm/models/Spark-X2.5.json

  "model_id": "XHToken/Spark-X2.5-4B",

• 源码证据：xinference/model/llm/transformers/spark_x2_5.py

  class SparkX25PytorchChatModel(_SparkX25PytorchMixin, PytorchChatModel):

• 源码证据：xinference/model/llm/vllm/patches/spark_x2_5_plugin.py

  def patch_spark_x2_5_plugin(env_path: str) -> bool:

核验提交：2c338fffdecb · 仓库更新时间：2026-10-27T13:30:56Z

• 源码证据：preflight/Modelfile

  FROM ./runtime/Spark-X2.6-4B-Q4_K_M.gguf

• 源码证据：noticeclear.py

  MODEL = 'spark-x26-preflight:q4km'

• 源码证据：noticeclear.py

  response=requests.post(BASE_URL+'/api/chat'

• 源码证据：noticeclear.py

  def export_copy(record,source,draft):

核验提交：f4a465796f48 · 仓库更新时间：2026-10-28T15:10:42Z

• 源码证据：models/download_gguf.sh

  URL='https://huggingface.co/XHToken/Spark-X2.6-4B-GGUF/resolve/fe31f8d64890302cc3a1d365d700357b4b75ba77/Spark-X2.6-4B-Q4_K_M.gguf'

• 源码证据：models/Modelfile

  FROM ./Spark-X2.6-4B-Q4_K_M.gguf

• 源码证据：app/model_client.py

  MODEL_NAME = os.environ.get("MODEL_NAME", "Spark-X2.6-4B")

• 源码证据：app/model_client.py

  payload["reasoning_effort"] = "none"

• 源码证据：app/server.py

  items.sort(key=lambda r: model_client.URGENCY_RANK.get(r["urgency"], 1))

核验提交：c86b2afe165e · 仓库更新时间：2026-10-30T23:25:56Z

• 源码证据：vllm_quantpack/__init__.py

  if spark.register(logger):

• 源码证据：vllm_quantpack/families/spark/config.py

  class Spark2_5Config(PretrainedConfig):

• 源码证据：vllm_quantpack/families/spark/model.py

  class Spark2_5ForCausalLM(nn.Module, SupportsPP):

• 源码证据：vllm_quantpack/families/spark/model.py

  def _split_fused_qkv(weights, config):

• 源码证据：vllm_quantpack/families/spark/model.py

  if tp != 1:

• 源码证据：vllm_quantpack/__init__.py

  patch_int4_kv(logger)

• 源码证据：vllm_quantpack/attention/int4_prefill.py

  MIN_QUERY = 64 # a chunk shorter than this is cheap on the Triton kernel

• 源码证据：vllm_quantpack/attention/int4_prefill.py

  MIN_CONTEXT = 8192 # below this the Triton kernel is already fast

• 源码证据：vllm_quantpack/attention/int4_kv.py

  def _step_lengths(cu_seqlens_q, seqused_k):

• 源码证据：vllm_quantpack/attention/int4_prefill.py

  def _cu_seqlens(length, device):

核验提交：6c53d2bb891b · 仓库更新时间：2026-10-31T14:46:06Z

• 源码证据：spark_eagle3/model.py

  class Eagle3Head(nn.Module):

• 源码证据：spark_eagle3/extract.py

  ap.add_argument("--layers", default="2,18,33")

• 源码证据：spark_eagle3/train.py

  opt.step()

• 源码证据：spark_eagle3/export_gguf.py

  w = GGUFWriter(str(out), "eagle3")

• 源码证据：scripts/train_s2b.py

  def compute_loss_mix(model, emb, d2t_t, t2d_t, feats, tok, lab, lam, mix_rate):

• 源码证据：spark_eagle3/extract.py

  def extract_contiguous_into(shim: Shim, store: FeatureStore, sources, *, n_ctx: int,

• 源码证据：scripts/train_s2b.py

  if args.answer_dir:

• 源码证据：spark_eagle3/corpus.py

  def iter_nlp_docs(paths, budget_chars: int, skip_chars: int = 0,

核验提交：1a2cedace485 · 仓库更新时间：2026-10-31T03:14:57Z

• 源码证据：convert/export_onnx.py

  SRC = Path("hf/Spark-X2.6-1.7B") # original checkpoint; tests/ and tools/ take both paths from here

• 源码证据：web/lib/spark.js

  executionProviders: ["webgpu"],

核验提交：af3a1e803d45 · 仓库更新时间：2026-10-31T03:14:55Z

• 源码证据：convert/export_onnx.py

  SRC = Path("hf/Spark-X2.6-4B")

• 源码证据：web/lib/spark.js

  executionProviders: ["webgpu"],

核验提交：8752e3389b2a · 仓库更新时间：2026-10-31T14:31:33Z

• 源码证据：rfl/providers.py

  MAAS_MODEL = "spark-x2.6-4b"

• 源码证据：rfl/providers.py

  out = self.model.generate(**enc, do_sample=False, max_new_tokens=self.max_new_tokens, pad_token_id=self.tok.pad_token_id)

• 源码证据：rfl/contract_v2.py

  def _legal_hits(s):

• 源码证据：mvp/review_flow.py

  def build_cards(record):

• 源码证据：scripts/score_locked_test.py

  def main():

核验提交：0ca79d68259f · 仓库更新时间：2026-10-30T10:15:23Z

• 源码证据：android/app/src/main/java/com/blackbox/diary/inference/LocalModels.kt

  val smeFile=File(root,"Spark-X2.5-4B-Q4_K_M.gguf")

• 源码证据：android/app/src/main/java/com/blackbox/diary/inference/AiRuntime.kt

  val raw=SparkSmeBridge.generate(graph.models.smeFile.path,prompt,cap)

核验提交：16714c599868 · 仓库更新时间：2026-10-30T23:35:01Z

• 源码证据：spark_npu/engine.py

  if "QNNExecutionProvider" not in session.get_providers():

• 源码证据：models/model.json

  "capacity": 512,

核验提交：a89a433562cb · 仓库更新时间：2026-11-02T00:21:36Z

• 源码证据：server/core.js

  small: { id: "Spark-X2.6-4B", label: "Spark X2.6 4B", url: () => settings.smallURL, temperature: () => settings.temperature },

• 源码证据：server/llm.js

  const kwargs = effort === "instant" ? { enable_thinking: false } : { enable_thinking: true, reasoning_effort: effort === "medium" ? "low" : "high" };

核验提交：4001b15fc2ce · 仓库更新时间：2026-11-01T09:45:42Z

• 源码证据：models/run-spark26.sh

  MODEL=${MODEL:-$ROOT/models/spark26-4b-Q6_K.gguf}

• 源码证据：models/run-spark26-dspark.sh

  --spec-type draft-dspark \

核验提交：f4bf872b3e92 · 仓库更新时间：2026-11-02T01:31:05Z

• 源码证据：configs/paths.example.toml

  model = '<drive>:\<folder>\Spark-X2.6-4B-SQL-Cypher-Agent-Q8_0.gguf'

• 源码证据：src/sqlrl/step2_pilots/rewards.py

  def reward(item: dict[str, Any], cfg: dict[str, Any]) -> float | None:

核验提交：4e3ecaf2a509 · 仓库更新时间：2026-11-01T22:17:49Z

• 源码证据：spark_work/convert_spark.py

  """Spark-X2.5 (XHToken, model_type spark2_5, remote code, THINKING) -> .litertlm.

• 源码证据：spark_work/x26_build.sh

  dl XHToken/Spark-X2.6-1.7B f18c032c7675f807c0fdffe281dbe124edba2605 $SM/Spark-X2.6-1.7B

• 源码证据：spark_work/x26_build.sh

  dl XHToken/Spark-X2.6-4B d1b5f26071101c27fcd07728e7a7770e428281f4 $SM/Spark-X2.6-4B