Scion-35B-A3B has 34.894 billion total parameters. The selected PQ2_0 experts + Q8_0 GGUF file is 11.337 GB.
Data snapshot: . 2 published tests in this catalog.
Model file and details
Provider
SkyIsNotGreen
Total parameters
34,893,918,848
Exact file size
11,336,908,288 bytes
Quantization / precision
PQ2_0 experts + Q8_0
Weight repository
SkyIsNotGreen/Scion-35B-A3B
Original model repository
SkyIsNotGreen/Scion-35B-A3B
Weight revision
ac1572d235f9a3ca2afb36e1f5f6219541303568
Model revision
ac1572d235f9a3ca2afb36e1f5f6219541303568
Runtime notes
Confirm model architecture and quantization support in your runner.
Recorded release date
2026-09-27
Date evidence
The repository was created on 27 September 2026. The pinned 16 October revision identifies the 34,893,918,848-parameter text model, exact 11,336,908,288-byte single-file body, Apache-2.0 license, custom PQ2_0 container, required runtime fork, sampled retention protocol and configuration-specific RX 7900 XT measurements. The post-boundary revision also adds an optional 66 KB correctness head; it is not included in the model byte count or ranking scores.
These scores were compiled from the linked sources. These published scores were not independently tested here. Different prompts, precision, thinking budgets and tools can change results.
Scion retention · HellaSwag 400
79 / 100
Publisher-run zero-shot HellaSwag accuracy on a fixed 400-task sample for Scion and same-base comparison quants. This is a limited retention panel, not the full benchmark.
Benchmark efficiency: 6.968 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Publisher-run zero-shot HellaSwag on 400 tasks. Scion and the BF16/Q4/IQ2/Q2 comparison rows used the same community protocol; BF16 and comparison quants were measured in one H100 session, while this release was measured locally and cross-checked against the pod.
Publisher-reported, not independently reproduced. A 400-task sample has about a +/-2 percentage-point confidence interval and is isolated from full HellaSwag. The selected file uses a custom PQ2_0 type that stock llama.cpp, Ollama and LM Studio cannot load; the publisher's moe-corr-runtime fork is required. CPU and ROCm/gfx1100 are verified, CUDA binaries are published but not end-to-end verified, Metal is untested and Vulkan is unsupported. The 11.34 GB file size excludes runtime, KV-cache and context headroom. A separate Strata run reported 118.0-118.2 tokens/s on one 20 GB RX 7900 XT at 4096 context, greedy decoding and 128 generated tokens; that measurement is configuration-specific and is not generalized to other hardware.
Publisher-run zero-shot Winogrande accuracy on a fixed 400-task sample for Scion and same-base comparison quants. This is a limited retention panel, not the full benchmark.
Benchmark efficiency: 6.726 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Publisher-run zero-shot Winogrande on 400 tasks. Scion and the BF16/Q4/IQ2/Q2 comparison rows used the same community protocol; BF16 and comparison quants were measured in one H100 session, while this release was measured locally and cross-checked against the pod.
Publisher-reported, not independently reproduced. A 400-task sample has about a +/-2 percentage-point confidence interval and is isolated from full Winogrande. Sub-1 percentage-point differences are treated as ties by the publisher. The exact selected artifact is text-only despite its base lineage, and its custom container requires the publisher's fork. Optional MTP and decision-head files are separate utilities and are not included in this score or the selected body size.
Context and version evidence: The repository and file name identify the selected model. Revision hashes are shown above when recorded. Tested context length, full hardware configuration and run date are not consistently recorded in this snapshot. Only settings explicitly stated in each benchmark record are known.
The 11.337 GB figure covers model weights in this file. It excludes the context cache, runtime, operating system and any separate multimodal files. It is not a measured RAM requirement or a token-generation speed.
These complete configurations pass the initial weight-size check for the selected 11.34 GB file. Allow extra memory for context and software. Check the model’s runtime notes and your inference software before buying.
Listing evidence: 2026-09-25. Amazon.com base price visible in an owner-supplied screenshot reviewed 25 Sep 2026; capture time not supplied. Dated reference, not a live quote. Excludes card signup offers, coupons, tax and shipping.
Listing evidence: 2026-10-03. Apple US launch price announced 22 September 2026: USD 1,699. Dated manufacturer reference, not an Amazon offer; tax and shipping may differ.
Listing evidence: 2026-09-24. GMKtec, EVO-X2 / 64GB / 1TB, recorded 22 Sep 2026; exact ASIN capacity identified 24 Sep. This is a dated reference, not a current Amazon price; tax, shipping and bundle contents may differ.
Memory matches are estimates, not tested compatibility or speed. As an Amazon Associate I earn from qualifying purchases.
Confirm model architecture and quantization support in your runner. The selected weights use 11.337 GB; this excludes context cache, runtime buffers and other programs.
These are the first listed tiers that pass our reserve policy, not measured minimum requirements. System RAM and VRAM remain separate. Use the memory tool to change the reserve for your context and runner.