jedisct1/Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP

Qwen3.8 Flash Next Uncensored oQ4e for oMLX

This is an oMLX-native quantization of `orcarouter/Qwen3.8-Flash-Next-Uncensored`. It was built from pinned revision fdf5fe3fd994c0cd6023c701d28f8820054f5aae. It targets a 128 GiB Apple Silicon Mac and preserves the Qwen4 vision tower, the checkpoint's own MTP head, and the checkpoint's native 262,144-token architecture setting.

The artifact is intended for oMLX and its OpenAI-compatible API. It can be used by coding agents and other clients that support that API.

Quantization

The default layout is oQ4e affine quantization with group size 64. The large PLE table uses four bits with group size 32 so oMLX can use its SSD-backed Qwen4 PLE path.

Strict oQe calibration covered 937 modules over 1,024 sequences of 512 tokens. Seven routed experts in base layers 0 and 1 were not selected by that fixed calibration corpus. All three routed projections in both affected blocks are therefore Q8 and do not use invented importance values. On the exact affected BF16 rows, Q8 reduced normalized reconstruction error by about 164 times compared with Q4.

The base language-model head, both base hyper-connection mixer projections, and the Qwen4 MTP embedding and hidden fusion projections remain BF16. These additions favor output quality and MTP draft acceptance over the last few GiB of compression.

The weights occupy 108,816,636,376 bytes across 22 safetensors shards.

oMLX setup

Add the downloaded model folder to oMLX and use a 100,000-token context limit on a 128 GiB Mac. Keep Qwen4 PLE SSD offload enabled for the full context window. MTP is included, but keep it disabled for coding agents. The matched tool soak was less accurate with MTP enabled.

Use these per-model settings as the starting point:

{
  "max_context_window": 100000,
  "qwen4_ple_ssd_offload": true,
  "mtp_enabled": false
}

For coding agents, set reasoning_effort to medium in the client profile and provider request body so a global setting cannot override the model profile:

reasoning_effort = "medium"
extra_body = { reasoning_effort = "medium" }

The model's prompt plus requested output must remain within 100,000 tokens.

Prompt caching

Prompt caching is recommended for coding agents and other clients that reuse long prefixes. Enable oMLX's paged SSD prefix cache, use a fast local SSD, and leave the in-memory hot cache disabled on a 128 GiB machine. The tested cache settings were a 100 GB SSD limit and a zero-byte hot cache.

The equivalent server options are:

omlx serve \
  --paged-ssd-cache-dir <fast-ssd-path>/omlx-cache \
  --paged-ssd-cache-max-size 100GB \
  --hot-cache-max-size 0

Keep max_context_window at 100,000 when caching is enabled. Caching reduces repeated-prefix prefill time, but it does not expand the safe token budget. The prompt plus requested output still must not exceed 100,000 tokens. Keep the balanced memory guard enabled because the first request must populate the cache and uses more memory than a cache hit.

Why 100,000 tokens

The checkpoint retains its native 262,144-token architecture setting, but that number is not a safe operating limit for this 128 GiB machine. The advertised 100,000-token limit is the largest round target that passed both exact API accounting and full-window retrieval with useful memory headroom.

At 100,000 tokens, the retrieval run peaked at 84.62 GiB of MLX memory and 91.01 GiB of physical memory, retained about 12.98 GiB of approximate available memory, and caused no swap growth. The 131,072-token attempt reached 118,784 processed tokens before oMLX predicted a 98.02 GB peak against its 96.77 GB balanced prefill safety cap. It could only fall back to 106,496 tokens, which was too close to the required memory reserve for a reliable advertised limit. Rounding down to 100,000 leaves room for oMLX, the client harness, and normal operating variance without changing the model's native positional metadata.

Prompt caching does not make 128K reliable on this configuration. A confirmed 98,304-token SSD cache hit inside a 100,000-token retrieval budget completed in 11.93 seconds, peaked at 89.58 GiB of physical memory, and caused no swap growth. The cache-populating request peaked at 96.23 GiB of physical memory and also caused no swap growth. At a 131,072-token budget, progressive cached prefill stopped at 104,448 and 118,784 processed tokens under the balanced guard and reached a 100.77 GiB physical peak.

Validation

The artifact has passed structural verification of every shard and tensor, the Q8 and BF16 safety policy, complete vision and MTP inventories, and the native context configuration. SHA256SUMS binds all 22 weight shards.

Testing with oMLX 0.6.3 on an M5 Max with 128 GiB passed an exact 100,000-token API budget. A 99,999-token uncached prompt plus one output token completed successfully, while both one-token-over requests were rejected. Peak MLX allocation was 84.62 GiB, peak physical memory was 90.82 GiB, and swap did not grow.

The long-context retrieval test also passed with a 99,744-token prompt and a 256-token output reserve. It recovered exact codes near token positions 569, 51,721, and 98,256, finished after 184 output tokens, peaked at 91.01 GiB of physical memory, and caused no swap growth from its stable baseline.

With the paged SSD prefix cache enabled, an exact repeat reused 98,304 of the 99,744 prompt tokens and recovered the same three codes. It finished in 11.93 seconds, peaked at 81.87 GiB of MLX memory and 89.58 GiB of physical memory, retained about 16.42 GiB of approximate available memory, and caused no swap growth.

The 131,072-token target is not supported by this safety profile. It reached 118,784 processed tokens before the balanced guard predicted a 98.02 GB peak against its 96.77 GB prefill cap.

The real-image test identified Half Dome correctly. MTP is operational and accepted 920 of 1,140 drafted tokens in the matched greedy suite, but exact visible parity held for only one of eight prompts.

Tool calling was evaluated with Swival as the harness. The quant is not tied to that harness, and other OpenAI-compatible coding agents and clients can use the same oMLX endpoint.

The 60-case evaluation meets the excellent tool-calling requirement with MTP disabled. All 50 positive cases passed on the first attempt and after recovery, and all 10 adversarial guard cases were protected. As with any tool-calling model, exact results also depend on the client's tool schema, parser, retry policy, and duplicate-call protection.

MTP remains operational, but the matched MTP-on matrix scored 48 of 50 positive cases and 10 of 10 adversarial guards. It accepted 9,765 of 11,150 drafted tokens across 139 model requests, but the two exact tool-argument regressions fail the no-regression requirement. Keep MTP disabled for coding agents.

License

The pinned source repository contains the Qwen Community License 1.0, which is preserved in LICENSE and includes additional commercial terms. The source model card labels the repository Apache-2.0, but that label does not match the license file in the pinned source revision. Review the included license before using or distributing this artifact.