Apex Flash 1 — DS4 mixed Q2 GGUF

Text-only conversion of Cantina Security Apex Flash 1, revision 28c647a6bd0444973a7ef3d940c24c67e64f52fe.

Converted directly from BF16 using IQ2_XXS routed gate/up, Q2_K routed down and higher-precision non-expert tensors. Q2 names a mixed recipe, not a uniform tensor format. No activation-calibrated importance matrix was used; IQ2 uses a weight-column-energy fallback. MTP tensors are retained, but speculative decoding is not qualified. Vision weights are omitted.

File: apex-flash-1-DS4-Q2.gguf, 96,505,818,912 bytes (89.88 GiB), 1,412 tensors.

Compatibility and verification

Requires a DS4 runtime supporting the GLM architecture and these mixed expert types. GGUF alone does not establish compatibility with llama.cpp, Ollama, MLX, LM Studio or their embedded engines. Memory needs depend on the exact runtime, expert cache and context; file size alone is not a RAM requirement.

Full output-payload integrity verification and the canonical source-equivalence audit passed. The conversion does not inherit the upstream checkpoint's scores.

Local qualification

Experimental, not a stable-agent release. Tested on 2026-10-04 using an Apple M5 Max MacBook Pro (Mac17,6), 128 GiB RAM, macOS 27.0.1 and the pinned DS4 runtime below. Context was 8,192 tokens; expert-cache target 16 GiB, zero full resident layers, SSD streaming, no MTP or draft model. The native allocation plan was 27.26 GiB; this is not a measured total-RAM requirement.

Throughput used 535 input / 128 output tokens, thinking off, temperature 0, top-p 1, top-k 0, seed 36 and zero cached prompt tokens. One tiny warm-up preceded the samples; the OS file cache was not cleared. This is a sequential pilot, not an ABBA comparison or a publisher benchmark.

Reasoning with thinking enabled completed correctly; state replay also passed. Quality grades check extracted answers/correctness, not JSON-only compliance; state replay appended prose despite the JSON-only request. The other cases failed valid-JSON output (None rather than null) and normal completion within the 256-token output cap. Four array-delimiter tool failures occurred in streaming and nonstreaming: Chat Completions altered delimiter text; Responses altered delimiter text and returned a string instead of the required array. Anthropic passed 12/12. These are exact-content/type tests, not security scores. Actual Prime/OpenCode tests hit harness usage-accounting/alias/receipt issues; agent delivery and executed coding are unqualified, not proven incompatible. The native server shut down cleanly and artifact identities were rechecked. An earlier 90-second harness socket timeout is retained as failed evidence; the completed native suite used a 600-second socket limit with the same rubric.

Larger contexts, vision, speculation and quality parity with BF16 are untested. Do not infer general coding ability or refusal behavior from these small tests.

Run on Apple Silicon

Use antirez/ds4 at 4bd088c20da0b905771fb94142cf12d669103c20, plus the included runtime patch. From this downloaded model folder, with Xcode Command Line Tools installed:

git clone https://github.com/antirez/ds4.git ds4-apex
cd ds4-apex
git checkout --detach 4bd088c20da0b905771fb94142cf12d669103c20
git apply --check ../runtime/schema-aware-glm-v2.patch
git apply ../runtime/schema-aware-glm-v2.patch
make -j4 ds4-server CC=clang
./ds4-server --metal -m ../apex-flash-1-DS4-Q2.gguf \
  --ctx 8192 --tokens 32768 --host 127.0.0.1 --port 18890 \
  --ssd-streaming --ssd-streaming-cold \
  --ssd-streaming-cache-experts 16GB --ssd-streaming-full-layers 0

Keep DS4_QUALIFICATION_RAW_CAPTURE undefined. Patch SHA256: 2ba7e633e0be1595d82b0b8191cc73c66292fbcf046b1b663e5d18fb88bed992. Expected patched ds4_server.c SHA256: bffea99f2bd5d122109cd64535a1b1781f8872dbbcd785420a1455d7cc6bff7f. Binary hashes vary with SDK/compiler; reproducing source is not a new live pass.

The API is http://127.0.0.1:18890/v1. It advertises glm-5.3-flash, its architecture alias, not the identity of the Apex fine-tune. Load only one large model at a time, retain native memory guards, and allow long request timeouts for reasoning. GUI integrations are not qualified by this upload.

provenance.json records pinned sources, recipe, verification scope, measured results and the complete GGUF SHA256. Runtime code has its own license in runtime/LICENSE.

Credits and license

Cantina Security and Yeta for Apex; Z.AI for GLM-5.3-Flash. Distributed under the upstream MIT license; see LICENSE.