werk1112

Model support audit — 2026-10-13

This audit checks the exact five requested products against vendor sources. An upstream release is not proof of executed Werk inference. All new executable paths remain IMPLEMENTED_UNTESTED until a real model/API smoke test succeeds; mock tests validate adapters, not model quality or account access.

Repository architecture and regression boundary

Werk is a Rust CLI/server, with canonical model manifests, task/capability metadata, runtime planning, generation adapters and a separate media service. src/backend/mod.rs inventories implemented runtimes: persistent llama.cpp server plus optional legacy bindings; Candle 0.9.2; managed/external vLLM; MLX/MLX-VLM/oMLX; ONNX Runtime; a Transformers compatibility adapter; experimental Burn; and Python media/text-analysis companions. TensorRT, OpenVINO, TensorFlow and CoreML format recognition is not evidence of a universal executable backend. ComfyUI and Automatic1111 are client integrations/API compatibility surfaces, not five additional native model engines. There was no general hosted-vendor provider abstraction before this change.

Existing FLUX.1 Schnell flow:

1. werk image generate / canonical werk run, or /v1/images/generations.
2. Shared InferenceService resolves the manifest, parameters and runtime plan.
3. CompanionMediaBackend sends the execution to its resident CompanionClient.
4. runtime/werk_media_companion.py::load_diffusers_pipeline selects diffusers.AutoPipelineForText2Image.from_pretrained(..., local_files_only=True). Diffusers resolves the checkpoint’s FluxPipeline configuration. There is no separate native Candle FLUX implementation in this repository.
5. The companion caches the configured pipeline, generates an image and writes an output artifact; the Rust output store supplies bytes/base64/URLs.

The FLUX 3 adapters do not replace or redirect that pipeline. A dedicated Schnell loader regression fixture and the existing companion/media tests cover this boundary. A real Schnell generation was not repeated in this environment.

Verified upstream facts and proposed/implemented routes

Product / exact identity	Input → output; capabilities	Weights, license, runtime and hardware	Official API / decision
Aleph Alpha Aleph-Alpha/Kolibri-1	Text → text; reasoning, tools, streaming through vLLM. German/English tokenizer and native chat template. 1,048,576 advertised context; ≤262,144 recommended.	78B total / 3.46B active, FP8 block-quantized MoE. Apache-2.0. kolibri1 / Kolibri1ForCausalLM; aleph-alpha-inference>=1 supplies matching vLLM/plugin/parsers. About 78 GB weights; vendor minimum includes 2×A100 80GB or 1×H200.	Extend existing vLLM path. No official hosted model endpoint verified. No GGUF/MLX/Candle/ONNX compatibility claim. Model card, config, plugin.
Mistral Large 4, API mistral-large-4	Text/images → text; tools, JSON/schema output and SSE. Public preview; vendor page advertises 1M context. Account/deployment limits remain authoritative.	Product page: granular MoE, 1.05T total / 52B active, 1.6B vision encoder. Actual downloadable artifacts are pending, so checkpoint architecture, tokenizer files, weight license and local engine compatibility cannot be validated.	Implement hosted /v1/chat/completions, Bearer key, random_seed, structured options and tool messages. Model, API, release announcement.
FLUX 3 Action, black-forest-labs/flux-3-action-droid	Synchronized wrist/left/right RGB frames + instruction + 8 state values → 32×8 absolute joint/gripper commands. Robot action prediction, not image generation or chat.	7B world-action model; DROID root and optimized variants, separate SO-101 package. DROID checkpoint: FLUX Kommunity License v1.0; code has separate terms. Shared encoders include Qwen3-VL-4B. Native Linux/NVIDIA PyTorch/NATTEN environment, per-package saved config/normalization.	Implement resident DROID worker using FluxActionPolicy; typed JSON predictions only. No official hosted API verified. SO-101’s 42×6 history/degree contract is deliberately not treated as DROID. Setup and schemas, DROID card, checkpoint license.
FLUX 3 Image, flux-3-image	Prompt plus optional 1–10 references → image. Generation/editing/layout instructions, aspect-ratio enum and resolution tiers.	No publicly downloadable checkpoint or compatible local runtime verified for this exact image model. FLUX.1 Schnell architecture cannot be assumed compatible.	Implement https://api.bfl.ai/v1/flux-3-image, x-key, returned regional polling URL, moderation and binary download. No seed, width, height, negative prompt, output-format or mode field in this endpoint. Schema, guide, product.
Eleven v4, eleven_v4	Text/voice pairs → speech audio; dialogue, language, stability/similarity, seed, normalization and streaming.	Proprietary service; no public weights/local engine verified. Model overview advertises 90+ languages / 10,000 characters. Dialogue endpoint recommends ≤2,000 total characters and ≤10 voices; adapter enforces that conservative endpoint limit.	Implement /v1/text-to-dialogue and /stream, xi-api-key, validated codec/rate/bitrate. Models page directs v4 to dialogue; product example also shows TTS, so the adapter uses the documented dialogue contract. Models, dialogue, streaming, OpenAPI.

The official Mistral weight destination shows an upcoming release with an ETA of 2026-10-31, after this audit date. The older Mistral-Large-4.0-1T05-A52B URL redirects there. No conversion, sharding configuration, weight license or local support is invented for it.

Feature negotiation and risk controls

• Model identity, provider and physical storage are separate: a provider source has no local checkpoint format; local registrations keep their inventory.
• Text/tool messages use existing chat abstractions. Action outputs use a separate typed endpoint with explicit units, dimensions and executed:false.
• BFL/Eleven extensions expose their actual contracts instead of silently translating unsupported Diffusers/OpenAI parameters.
• New support status is independent of credentials configured, backend installed or preload success. Discovery never calls a paid generation endpoint.
• Kolibri uses the vendor parsers and native template; reasoning effort is forwarded to template kwargs. Its synthetic 32-token preload disables thinking so hidden reasoning alone cannot exhaust the short warmup.
• Startup preload preserves ordering, deduplication and retention; a failed target now stops server startup. DROID workers use the same pinned runtime-cache transaction; cloud preloads only prepare the HTTPS client and validate local credential syntax. They do not promise provider readiness or warmed weights.
• POST network/5xx failures are not automatically replayed: a generation may already have been billed. Explicit 429 rejection is retried with bounded backoff. Redirects are disabled; BFL result downloads never receive the key.

See implementation and test evidence and the status matrix with runnable commands.

This site is open source. Improve this page.