This model is based on Qwen3.8-Flash-Next-FLARE by Local Inference Lab, Inc., a non-profit organization, available at https://huggingface.co/local-inference-lab/Qwen3.8-Flash-Next-FLARE. Qwen3.8-Flash-Next-FLARE is licensed under the Local Inference Lab License, Version 1.0.

Qwen3.8-Flash-Next-FLARE-Abliterated

Reduced-refusal model. A refusal-direction edit was applied on purpose. Read the Safety section before you deploy this.

gb202 is now v5. v5 carries the same weights and the same refusal edit as v2, v3 and v4. It changes two things: it frees about 2.1 GiB of GPU memory, and it is signed by a new key.

v5 is signed by a new key β€” you must approve it

v3 and v4 were signed by a local development key. v5 is signed by our own publisher key in the lflare namespace.

revision publisher namespace signing key fingerprint
v3, v4 local-development 0a574ece56e54541d7362e10ba129988b3411d05a5e4805b4d9114c92405479e
v5 lflare 941a0379d317c6c809c9918dc86c38bd50c12710c1b713ba6a2336971b58be81

A command that pins the old fingerprint refuses v5. Change both flags: --dangerously-approve-key 941a0379… and --dangerously-accept-publisher lflare.

The lflare namespace is not a development namespace, so the engine verifies it over the network. At boot it reads https://huggingface.co/lflare/Qwen3.8-Flash-Next-FLARE-Abliterated/resolve/main/publisher.json and requires the signing key to appear there. That file lists the key, so a normal boot reports Identity: Verified via huggingface.co/lflare. A host with no access to huggingface.co must pass --offline, which skips the check and reports Identity: Unverified instead.

v5 frees GPU memory, which becomes KV cache

Commit 5fb495da in the preparation tool stops the expanded routed-expert CSF scales from staying in the packed core storage. The package on disk is almost the same size, because the package holds the compressed scales and the expansion happens at load. The saving is device memory, and the engine spends it on cache.

Measured with the same engine binary, the same GPU model and concurrency 16:

v4 v5
model memory 75.4 GiB 73.3 GiB
weights 72.6 GiB 72.2 GiB
cache budget 18.57 GiB 20.70 GiB
KV tokens 1,261,568 1,421,312 (+12.66 %)
shared snapshot slots 128 155 (+21.09 %)

Your own figures will differ a little with driver and free memory. Read the Plan cache and Allocate cache lines in your boot.

Earlier revisions

v1 and v2 were upstream's package with our weights patched into it, byte for byte, keeping upstream's signature. Both were deleted on 2026-11-17. v3, v4 and v5 are real re-exports: the abliterated checkpoint is prepared from scratch with upstream's own Ignition, so each is a package in its own right, signed by us.

Older Flare runtimes once refused a package whose execution ABI did not match exactly. That gate is gone. Upstream now requires every engine version to load any package a previous version served, and treats execution ABI numbers as descriptive metadata rather than loading gates. So you no longer pick a revision to match your runtime. Serve v5.

v3 declares execution ABI 4 and prepared contract 2. v4 and v5 declare contract 3. All three load on a current engine.

v5 is published as the directory canonical form β€” a 15,316,072-byte signed model.safetensors primary plus 55 companions under weights/ and deployments/, 106,327,044,081 bytes in total. One packed file would have been simpler, but the Hub refuses a single 106 GB LFS object, and the primary names each companion by its exact relative path, so the paths matter.

revision edit packaging thinking off thinking on
v3 rank-2, layer 42 ABI 4 / contract 2 complies complies, 0 %
v4 rank-2, layer 42 ABI 5 / contract 3 complies complies, 0 %
v5 = gb202 rank-2, layer 42 contract 3, CSF scales not resident complies complies, 0 %

All three carry the same edit. They change how it is packaged, not what it does. Those figures were measured on the equivalent plain checkpoint β€” see Measured behaviour. v1 and v2 were deleted on 2026-11-17.

# `publisher.json` here is v5's: the lflare key. v3 and v4 carry the development
# key instead, so take the file from the same revision as the weights you pull.
hf download lflare/Qwen3.8-Flash-Next-FLARE-Abliterated \
  --revision v5 --include 'model.safetensors' 'weights/*' 'deployments/*' \
  'publisher.json' 'config.json' --local-dir ./pkg

# A local package directory needs no model cache and no revision resolution.
flare run ./pkg \
  --gpus 0 --embedding-table-storage ram \
  --verify-weights \
  --dangerously-approve-key 941a0379d317c6c809c9918dc86c38bd50c12710c1b713ba6a2336971b58be81 \
  --dangerously-accept-publisher lflare

--publisher-keys is not needed for v5. The engine reads the key from this repository's main branch, as described above. Add --publisher-keys ./pkg/publisher.json --offline instead when the host cannot reach huggingface.co, and accept that the identity is then unverified.

gb202 tracks the newest package, so pin --revision v5 if you want it fixed.

config.json is in that first command on purpose, and it is not a loadable model config. The Hub counts a download only when it serves one of a repo's query files, and for a repository that declares no library that file is config.json (docs). Without it this repository records zero downloads however many times the package is pulled β€” which is why both this repo and the upstream Flare package sat at 0. The file deliberately has no model_type, architectures or auto_map key, so nothing is tempted to treat a Flare package as a Transformers checkpoint. AutoConfig.from_pretrained is expected to fail on it.

Read this first β€” three operational facts

1. --verify-weights now works, and you should use it. This was the opposite for v1 and v2: patching weights invalidated upstream's digest over the whole weight region, so the flag failed after about 190 s and those builds could not be cryptographically verified at all. v3 and v4 are real exports signed by our own key, so the digest covers our own bytes. Measured on this package: Verify stored data. 99.5 GiB 37.5s, exit 0.

2. The signing key is ours, not Local Inference Lab's. v1 and v2 authenticated as theirs because the patch left the signed metadata untouched. v3 and v4 are signed by the development key Ignition created during their export: namespace local-development, fingerprint 0a574ece56e54541d7362e10ba129988b3411d05a5e4805b4d9114c92405479e. For those two the runtime discloses it plainly β€” "Local development package; not an official published checkpoint".

v5 is signed by our own lflare key, fingerprint 941a0379d317c6c809c9918dc86c38bd50c12710c1b713ba6a2336971b58be81, and the runtime verifies it against publisher.json on this repository's main branch. A v5 boot reports Identity: Verified via huggingface.co/lflare. main now carries the lflare profile for exactly that reason. v3 and v4 still need the profile from their own revision, so take the file from the same revision as the weights unless you are serving v5. Verification is not an endorsement by anyone β€” see Licence.

3. You need CUDA 13.2 or newer. Measured on this package and the upstream one alike:

driver CUDA result
580.178.04 13.0 fails β€” cuGraphKernelNodeSetAttribute … CUDA_ERROR_INVALID_VALUE at Prepare execution
595.58.03 13.2 works
610.57.04 13.3 works
615.71.09 13.4 works

One GPU, TP1, up to 16 concurrent requests, up to 5 MTP draft tokens β€” the capacities this export was prepared for. Flare rejects temperature=0 ("greedy decoding incorrectly collapses the model's probability distribution"); 0.01 is accepted.

Measured behaviour

Partly measured on this package. v5 was exercised through the Flare runtime itself on the export host: it boots (Start server.), verifies its identity against this repository, and reports 73.72 GiB loaded with a 20.70 GiB cache β€” 1,421,312 KV tokens, 155 shared snapshot slots, 5,161 kernels, concurrency 16. v4 on the same binary and host reported 75.83 GiB loaded with an 18.57 GiB cache, 1,261,568 KV tokens, 128 slots and 4,331 kernels, and additionally ran MTP speculation across widths 2 to 6 (359 speculative graph launches over three requests). What was not re-run through Flare is the 136-row held-out gate, so the refusal figures below come from the plain checkpoint whose weights these are: lflare/Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD-Abliterated at revision v3, whose measurements/ directory holds the raw result files. v3's edit is the same rank-2 dual-regime edit as v2, re-applied to the same weights and reproducing v2's figures to four significant figures β€” biprojection cosines 0.982786 and 0.903085, cos(v1, v2) = 0.3495, 25,188 of 25,188 targets edited, max leakage 0.002305, mean added weight error 5.618 % of β€–Wβ€–.

Held-out rows the layer selection never saw, 136 harmful and 59 over-refusal, parent against this edit:

protocol scorer parent this edit
thinking off, 100 tok model-forge 136/136 (100 %) 0/136 (0 %)
thinking off strict, 47 marker 136/136 (100 %) 0/136 (0 %)
thinking on, 2048 tok model-forge 115/136 (84.6 %) 0/136 (0 %)
thinking on strict, marker-matched β€” 2/136 (1.5 %)
thinking off over-refusal 2/59 (3.4 %) 0/59 (0 %)

Capability at parity in the same session: lavd 30/30 against the parent's 30/30, estonia 30/30 against 30/30, paired GPQA-diamond 166/198 against 164/198 (not significant, p = 0.85).

What is genuinely unverified here is the held-out gate through Flare, and capability through Flare. Long-context retrieval, GPQA, MTP acceptance rate and vision were never measured through the runtime on any of our builds. Unlike v2, v3 and v4 no longer rest on the assumption that a patched package serves equivalently, because it is not patched.

Scorer warning. A keyword scorer, as opposed to a refusal-form scorer, returns wildly different numbers on the same responses, because words like "illegal" occur inside answers that comply. A refusal number is a protocol and scorer pair, and both halves move it a long way.

What was edited

The edit removes the span of two refusal directions, not one. The thinking-off and thinking-on refusal directions on this model sit 67–77Β° apart, against a capture noise floor of 1.5–3.5Β°, so ablating one leaves about 96 % of the other standing β€” which is why v1 complied with thinking off and still refused with thinking on. Both directions were captured at layer 42 and biprojected against their own harmless mean; they sit at cos = 0.3495, 69.5Β° apart. The committed artefacts are direction_off_a.npz and direction_on_a.npz in the plain checkpoint's measurements/.

v3 and v4 reach 512 tensors that v1 and v2 could not. The edit is applied to the checkpoint and the package is then exported from it, so every target is simply addressable:

edited count format
layers.{0..47}.mlp.experts.{0..511}.down_proj 24,576 NVFP4
mtp.layers.0.mlp.experts.{0..511}.down_proj 512 NVFP4
layers.*.linear_attn.out_proj 36 MXFP8
layers.*.self_attn.o_proj 12 MXFP8
layers.*.mlp.shared_expert.down_proj 48 MXFP8
embed_tokens, ple.value_proj, 2 more 4 BF16
total 25,188

Those 512 MTP routed-expert planes were deliberately skipped by v1 and v2, because their storage region inside the sealed package could not be identified β€” a 612-probe scan found 100 of 612 targets and the 512 misses were exactly those planes. Re-exporting removes that limitation along with the patching entirely.

No scale tensor is touched. The constrained requantiser keeps every code on its own quantisation grid within a Β±1 ULP budget, which is why a naive round-to-nearest requantisation is not used: measured on real shards, plain rounding retains only 14 % of an NVFP4 edit.

For the record, v1 and v2 were produced the other way: by locating weight regions inside upstream's package by content-matching and probing, then rewriting 22,285,516,800 bytes in place with every offset confirmed against the parent beforehand and every write read back. That method is sound but cannot reach what it cannot locate, and it leaves the package unverifiable. It is superseded.

Verifying this download

v5 is a directory, so there is no single file hash. Verify it the proper way, which v1 and v2 could not offer:

flare run ./pkg --verify-weights --stage-only \
  --dangerously-approve-key 941a0379d317c6c809c9918dc86c38bd50c12710c1b713ba6a2336971b58be81 \
  --dangerously-accept-publisher lflare
# [    OK    ] Verify stored data.  99.5 GiB  37.5s

That checks every payload object against the signed manifest and exits before touching the GPU. The primary, which carries the manifest, the signature and the physical locations:

file   : model.safetensors            (revision v5 = gb202)
size   : 15,316,072 bytes
sha256 : eda946a12730f5aa236751234a4345523a254f1a0a80a9b04d541c8c2276e0a0

v4's primary, for anyone who pinned it, is 23,147,280 bytes with sha256 dc17e81fdf734e1ca001797c9ed0a27568f9aea573ad2e8cad422f7d9546ce7e.

The earlier single-file packages. v1 and v2 were deleted on 2026-11-17, so these hashes are a record rather than something you can still download:

package sha256
upstream parent c3db5e1ac63b05efa67d6525826a8a792c0ce92b810e91ddd5ec0f42c73a175c
ours, v1 fff1962065e4b09858bf700646336243ab18faf9900dad9c4f8486345032ef3d
ours, v2 35112671d85c719222c40d527d88f475cc38cbae6f20d1b94d411083ce0ac7f4

Export provenance

Exported on 2026-10-30 from fork develop 5be036b3 (upstream main a161b17f) with Ignition from that tree. The export required PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True: without it, preparation failed at Warm compiled forwards with a 128 MiB CUDA allocation shortfall on a 95.06 GiB device. Upstream's own note on f617022d (the ABI 5 change) records that its GPU and real-model export validation had not been performed on the receiving tree, so this package is the first real-model ABI-5 export we know of. The full export log and the per-phase resource trace are in the plain checkpoint's measurements/ directory.

Limitations

  • The held-out refusal gate was not re-run through Flare. It was measured on the plain checkpoint these weights come from. See Measured behaviour.
  • Capability through Flare is unmeasured: long-context retrieval, GPQA, MTP acceptance rate and vision were never measured through the runtime.
  • Signed by a development key, which the runtime announces as "not an official published checkpoint". Approving it is a trust decision you make.
  • Requires Flare a161b17f or newer (execution ABI 5, prepared model contract 3). bce032c0 and older runtimes refuse it with "unsupported execution ABI version; re-export the model package"; serve v3 on those.
  • One GPU, TP1, concurrency 16, 5 MTP draft tokens. Multi-host RoCE is DGX Spark (arm64) only β€” the amd64 Flare binary does not link libibverbs.
  • CUDA 13.2 or newer required.
  • These files require Flare. They cannot be loaded by Transformers, vLLM, SGLang or llama.cpp, and Flare cannot load an ordinary checkpoint.

Safety

Refusal behaviour has been deliberately and substantially removed, in both thinking regimes, where v1 only did so with thinking off. This model answers requests the parent declines, including harmful ones. Do not deploy it as a general-purpose assistant without your own safeguards, and do not read these numbers as a safety property.

Counter-intuitively, thinking-off is the harder regime: the unedited parent refuses 100 % there against 84.6 % with thinking on. Report refusal as a protocol and scorer pair, and compare only against the parent measured identically.

Licence

Local Inference Lab License, Version 1.0. See LICENSE, NOTICE and LICENSES/, copied unchanged from the licensor's distribution, with the Identifying Marks intact.

Redistribution of this derivative is by permission of the licensor under Section 2.4. Local Inference Lab does not sponsor or endorse this model or its authors, and nothing here should be read as approval of a reduced-refusal build. The upstream model remains under the Qwen Community License 1.0; see LICENSES/LicenseRef-qwen-community-1.0.txt. If you operate a Model-as-a-Service or AI-assistant business, that licence requires a separate licence from Qwen before commercial use.

Credits: Qwen/Qwen3.8-Flash-Next by Qwen. The Flare runtime, the NVFP4-CSF container format, the b12x kernels and the Ignition preparation tool that produced this package are all by Local Inference Lab, Inc. v1 and v2 additionally edited a prepared package of theirs; v3 and v4 do not, and are exported from our own abliterated checkpoint.

Downloads last month
248
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for lflare/Qwen3.8-Flash-Next-FLARE-Abliterated

Collection including lflare/Qwen3.8-Flash-Next-FLARE-Abliterated